passepartout/docs/DESIGN_DECISIONS.org

• One single agent
• The Unified Memory Argument
• Org-Mode as Unified AST
• Homoiconicity as Foundation
• The Probabilistic-Deterministic Split
• The Dispatcher as Learning System
• The REPL as Cognitive Substrate
• Literate Programming as Discipline
• The Evaluation Harness
• Observability and the Thought Trace
• The MCP Strategy
• Local-First Architecture
• Token Economics and Performance Advantage
  • The Core Insight: LLM as Expensive Resource, Not Default Engine
  • Per-Task Type Analysis
    • Coding (debugging, refactoring, PR review)
    • Knowledge Management (Zettelkasten, research, note-taking)
    • Day-to-Day Life Management (calendar, tasks, reminders)
    • Chatting (casual conversation)
  • The Dispatcher Learning Curve: Cost Decreases Over Time
  • Local LLM Viability
  • Open Questions and Risks
  • Comparison Summary

This document captures the rationale behind key architectural choices. It is not a specification - it is a thinking medium for future architects and contributors who need to understand why the system is built this way, not just how.

One single agent

The AI industry has developed an intuition toward multi-agent systems as the default solution to hard problems. Multiple agents spawn, delegate, coordinate, debate, and consensus their way toward solutions. This pattern is compelling in demos and genuinely useful in specific contexts - but it has become a default assumption that warrants scrutiny.

When context windows grew expensive and task complexity increased, the response was natural: split the problem across agents, each handling a slice. But this architectural choice carries hidden costs that are rarely acknowledged in the enthusiasm of implementation.

The synchronization tax is the most immediate burden. Each agent operates with partial information, and maintaining coherence requires continuous state reconciliation. Tokens and processing cycles are spent not on the task itself, but on protocol overhead - who holds what, who decided what, who is correct when they disagree.

Fragmented context is the deeper problem. When Agent A writes a function and Agent B modifies a type it depends on, neither has the full picture. Integration failures emerge not from individual incompetence but from systemic communication gaps. Single-agent systems avoid this entirely: one brain holds the complete model, every decision is made with full visibility.

Audit trails become complex in multi-agent systems. A decision traced through a single-agent system has a clean, linear history. A decision traced through a multi-agent system branches and forks, with each agent's reasoning partially overlapping and partially conflicting.

None of this is to say multi-agent systems are never appropriate. Embarrassingly parallel workloads - scanning ten thousand files, processing batch jobs - benefit from parallelism regardless of context. When distinct expertises are required and cannot coexist in one model, delegation makes sense. In adversarial scenarios where conflicting goals are features, multi-agent architectures shine.

But the default assumption that complex reasoning tasks are best solved by multiple agents is unproven and likely wrong for the engineering domain. Claude Code is a single-agent system. It handles 50-file refactors, debugs complex stack traces, writes tests, and navigates large codebases. The assumption that you need five agents to do what one well-designed agent can do is an industry habit, not a technical necessity.

Passepartout is single-agent by default not from limitation but from conviction: for reasoning-heavy work where coherence matters, a unified memory space and single decision-making locus are architectural assets, not constraints.

The Unified Memory Argument

If single-agent architecture is the decision, unified memory becomes the mechanism that makes it viable. The critical question is not "how many agents" but "how does the agent manage context without saturating."

Context window limits are largely a symptom of lazy architecture. The default approach - stuff everything in, hope the model figures it out - works poorly at scale. A more principled approach inverts the problem: the system should hold effectively infinite context, with the active window kept lean through intelligent management.

Lazy loading is the core technique. When an agent needs information about a function, it does not load the entire codebase. It loads precisely what the function does. Context stays lean - 2,000 to 4,000 tokens - while the full context remains accessible through retrieval.

Compaction events are scheduled during idle cycles. The system extracts new facts from active context and writes them to permanent storage. Active context is wiped clean, not because space ran out, but because the information has been preserved in a form that can be retrieved when relevant.

Org-mode as externalized memory solves the persistence problem elegantly. Every decision, every note, every task lives in plain text files the user already owns. The agent does not maintain a separate database. It queries files it can already access, modifies files it already owns.

Retrieval is the key primitive. Semantic search across Org files finds relevant nodes. The agent does not hold the full context - it holds pointers to context, loaded on demand. This is how a single agent handles tasks that would saturate a naive multi-megabyte context window.

The unified memory argument is not that infinite context is free. It is that with proper architecture, effective infinite context is achievable without the synchronization and fragmentation costs of multi-agent systems.

Org-Mode as Unified AST

Passepartout makes a bet that most systems consider too expensive to place: that humans and machines should share the same file format. That bet is Org-mode.

Most systems separate human-readable notes from machine-readable data. The user writes Markdown. The system stores it, indexes it, searches it. But internally, the system maintains its own model - a database, an object store, a knowledge graph - that is disconnected from the Markdown. When the user dies or leaves, the Markdown survives but the model must be reconstructed.

Passepartout refuses this separation. The Org file is not a representation of the data. The Org file IS the data. The same text that the user reads and edits is what the system parses and operates on. org-element reads an Org buffer and returns a tree structure that is the direct Lisp representation of the file's content.

This has several profound implications.

First, there is no translation layer between human and machine. When the agent writes a skill, it writes Org text that is immediately readable by the human who owns the file. When the human writes a note, it is immediately accessible to the agent as a native data structure. The communication is not mediated by a schema or an import/export process.

Second, the format is genuinely readable by both parties, not just technically accessible. Org-mode's syntax is human-friendly: headlines begin with asterisks, properties live in drawers, tags are labels after colons. The human does not have to understand the full Org specification to read what the agent wrote. The agent does not have to handle edge cases in human notation.

Third, the format is stable across decades. Org-mode has been in active development since 2003. The files written today will be readable by Org-mode in 2040. There is no schema migration, no database upgrade, no vendor lock-in. The human's notes survive the system.

Fourth, the format is universally available. Org-mode is free software. The files are plain text. There is no proprietary format to decode, no application to purchase, no cloud service to access.

Fifth, the format is header-aware and sparse-tree capable. Org-mode's headline hierarchy is not just formatting - it is a semantic structure the system can query. The agent can retrieve only the relevant subtree under a heading, ignoring the rest of the file. This is fundamentally different from Markdown, where the entire file must be loaded or the retrieval logic must parse and filter at the string level.

Sparse tree retrieval is the key to efficient context management. When the agent needs information about the openctl-db function, it queries for the openctl-db subtree specifically. It receives exactly the code, documentation, and metadata under that heading - nothing more. The context stays lean not because the file was pre-split but because the retrieval is structural. In a Markdown system, the agent either loads the entire file (expensive, noisy) or relies on imprecise grep-like search (fragile, loses hierarchy). In Org-mode, retrieval is precise, hierarchical, and cheap. The heading boundary is the access boundary.

Sixth, Org-mode unifies what every other format fragments. A single Org file contains the headline hierarchy, prose documentation, source code blocks with live evaluation, tags for categorization, metadata in property drawers, TODO state for task management, timestamps and deadlines, and links to other nodes. Markdown cannot express TODO state without external tools. JSON cannot contain prose. YAML cannot embed runnable code. Each format serves one purpose; Org-mode serves all of them. When the agent reads a skill file, it reads documentation, code, dependencies, metadata, and task state in one parseable structure. When the human reads the same file, they see the same information rendered in a human-friendly form. No other format achieves this unification without maintaining parallel files or external databases.

Seventh, a skill lives in one Org file, not a directory. The standard pattern for a software project is a directory containing README.md, package.json, src/main.py, src/utils.py, tests/test_main.py, scripts/deploy.sh, and config.yaml. Each file type is isolated by convention: prose lives in README, code lives in src, tests in tests, configuration in config. This fragmentation means the skill is not a single object the system can reason about - it is a collection of files the system must assemble. Passepartout's skills violate this convention deliberately. Each skill is one Org file. The file contains the skill's documentation, the skill's code, the skill's metadata, the skill's TODO state, and the skill's dependencies on other skills. There is no directory to navigate, no external files to locate, no risk that the README describes behavior that the code does not implement. The skill is a single atomic unit: readable by human and machine, editable by both, versionable as one entity.

The unified format is what makes the memory architecture work. The agent's memory is not a database that the user cannot inspect. It is a folder of Org files that the user can read, edit, and understand. The agent manipulates these files directly, using the same tools the user would use. There is no hidden state, no shadow database, no model that differs from the source.

This is what "sovereignty" means in technical terms: the user owns the data in a format they can access, and the agent operates on the data in the same format they own.

Homoiconicity as Foundation

Common Lisp is homoiconic: code and data share the same representation. A Lisp program is a list, and a list is a Lisp program. This is usually presented as a curiosity, an interesting property that enables macros. In Passepartout, it is the foundational enabling property of the entire self-modification architecture.

When code is data, the agent can read its own source the same way it reads a text file or an Org buffer. There is no AST parser required, no external tool to extract the function object from the running image. The agent evaluates (read-from-string source) and the result is executable Lisp. The representation it manipulates is the same representation that the runtime executes.

This is not true of most languages. In Python, the agent can inspect an AST through the ast module, but that AST is a foreign object - a data structure that represents code but is not code itself. The agent can see that a function takes certain arguments and returns a certain type, but it cannot treat the AST as a live object it can modify and re-evaluate. In C, the agent cannot inspect its own compiled machine code at all.

In Lisp, the distinction between code and data is a convention, not a barrier. The agent's skills are lists. The agent can take a skill, extract a function definition, modify the body, wrap it in a new list, and evaluate it. The modification is surgical: it changes exactly what it intends to change, with no risk of corrupting adjacent state, because the representation is a tree that the runtime understands natively.

Runtime introspection is therefore native. The agent does not need a debugger API or a reflection protocol. It operates on its own code as data because its own code is data. (describe 'function-name) returns the function's documentation. (function-lambda-list 'function-name) returns its parameters. (macroexpand-1 '(defskill …)) shows what the macro produces. There is no impedance mismatch between the agent's reasoning and the system's representation.

Self-modification is the practical consequence. The agent can detect an error, locate the erroneous function, generate a corrected version, and hot-reload it into the running image. The correction is not applied to a file that requires a restart - it is applied to the live object that the system is currently executing. This is what makes the self-editing skill viable: the agent can fix itself without stopping.

In v3.0.0, when the symbolic engine takes over the reasoning core, homoiconicity becomes the bridge between the neural and symbolic layers. The neural engine generates proposals as s-expressions. The symbolic engine evaluates them against formal constraints. The result is a modification that is simultaneously a data structure the symbolic engine can analyze and code the runtime can execute. The two representations are identical by construction.

This is the technical meaning of "Lisp as Governor": not just that Lisp orchestrates the other components, but that the representation of the system is uniform and inspectable at every level. There is no hidden state, no opaque machine code, no representation that the agent cannot reach into and modify. The system is legible to itself by design.

Self-Modification Without Boundaries

Other systems that support self-editing draw a line between the core and the skills. Hermes can modify its skills at runtime, but the core harness is protected - editing it requires a restart because the core is treated as privileged code that cannot be safely modified while running.

Passepartout has no such boundary. The "thin harness, fat skills" distinction describes where complexity lives, not where authority flows. The harness is small by design, but it is not privileged. The agent can read and write any part of the system - including the very code that is currently executing - without restarting.

This is only possible because Lisp code is mutable data at runtime. In a compiled language, the machine code for a running function is locked in memory, protected by the call stack, impossible to modify safely. In Lisp, the function object is a list you can modify with setf. When the agent changes a harness function, the running image immediately reflects the change. The next invocation uses the new code. There is no restart, no special boot mode, no distinction between development and production.

The implications extend beyond convenience. A system that cannot modify its own core is a system that has limits on its own adaptability. It can learn skills but not improve its own structure. It can grow but not evolve. Passepartout's lack of a core boundary means the system can improve its own reasoning engine, fix bugs in its own cognition, and evolve its own architecture - all while continuing to operate.

This is the final expression of homoiconicity: not just that code is readable as data, or that skills are modifiable, but that the entire system - including the parts that other systems protect - is open to modification. There is no ceiling on self-improvement. The agent can rewrite the very code that rewrites itself.

Lisp and the AI Dream

Lisp was invented in 1958 by John McCarthy with artificial intelligence explicitly in mind. Its design - code as data, runtime mutation, symbols and lists as first-class constructs - was shaped by the belief that a truly intelligent machine would need to reason about and modify its own reasoning. For decades, Lisp machines were the closest thing to thinking machines that existed.

Then the AI winter came. Symbolic AI fell out of favor. Statistical learning and neural networks dominated. Lisp was relegated to niche applications and academic curiosity. The machine that was designed for AI was never used for the task it was designed for.

Six decades later, neural networks have arrived at the problem from a different direction. They can learn and generalize, but they hallucinate, cannot explain their reasoning, and cannot safely modify themselves. The neuro-symbolic synthesis - combining neural pattern recognition with symbolic reasoning - is recognized as the path toward AI that is both powerful and trustworthy.

Lisp's time may finally have come. Not as a replacement for neural networks, but as the governor that makes them safe - the symbolic engine that verifies what the neural engine proposes, the homoiconic substrate that allows the system to inspect, modify, and improve its own reasoning. The machine that was designed for AI in 1958 may be the exact machine needed for AI in 2026 and beyond.

The Probabilistic-Deterministic Split

The architecture divides cognition into two fundamentally different reasoning systems. This is not arbitrary engineering but a structural response to a fundamental truth: probabilistic systems will hallucinate, and you cannot build reliable autonomy on an unreliable foundation.

An LLM is a statistical engine. It generates outputs based on patterns in training data. It is remarkable at translation, generation, pattern matching, and fuzzy reasoning. It can take messy human intent and produce structured queries. It can take structured results and produce natural language. It is, in the terminology of the system, the creative brain.

But it cannot be trusted. Not because it is poorly designed or insufficiently trained, but because hallucination is a fundamental property of probabilistic inference. The model generates the most likely continuation, not the correct one. Given sufficient context, the most likely continuation is correct. Given novel context, it is often wrong in confident-sounding ways.

The deterministic engine addresses this by being what the probabilistic engine is not: mathematically rigorous, formally verifiable, and incapable of hallucination by design. It operates on explicit symbolic representations - lists, property lists, knowledge graphs - not on floating-point activations. When it evaluates a path confinement check, it returns true or false, not a probability distribution.

The division of labor is architectural. The LLM handles the fuzzy interface between human language and structured representation. It translates what the user wants into what the system can reason about. The deterministic engine receives those structured representations and evaluates them against formal invariants. It decides whether to execute, not whether the translation was semantically plausible.

This separation is the source of Passepartout's safety guarantee. Other agents add "guardrails" as an afterthought - a layer of filtering around a dangerous core. Passepartout makes the division explicit: the LLM never touches the file system, never executes a command, never modifies memory. It generates proposals. The deterministic engine evaluates and executes. The dangerous operations are never in the probabilistic path.

The split also explains why the system gets safer over time without the LLM improving. The deterministic engine accumulates rules. The LLM proposes actions, the engine evaluates them against a growing rule set. Early versions block obvious dangers. Later versions block sophisticated attacks that were previously unknown. The safety grows logarithmically with the number of interactions, not linearly with model capability.

The Dispatcher as Learning System

The Dispatcher begins as a static guard - a set of rules that block obviously dangerous actions. But defining "obviously" is the hard problem. The agent encounters situations the rules do not anticipate. The Dispatcher must grow.

The human-in-the-loop exception is the seed. When the LLM proposes an action the Dispatcher does not recognize, the system does not default to blocking or allowing. It suspends. It writes the proposed action to an Org buffer in a format the human can read and understand. The human reviews and approves or denies. The Dispatcher observes the decision.

From this single observation, the Dispatcher extracts a rule. Not merely "allow this specific action" but "allow this class of actions parameterized by these dimensions." The human approved a write to ~/projects/myapp/src/core.clj. The Dispatcher generalizes: writes to ~/projects/*/src/*.lisp are approved for this session, or for this project, or indefinitely depending on the context and the user's pattern of decisions.

Shadow mode is where rules are tested before deployment. When the Dispatcher encounters a novel situation and is uncertain, it can run the proposed action in a simulated environment. It observes the side effects - what files would be modified, what processes would be spawned, what network calls would be made. If the simulation produces dangerous side effects, the rule is discarded. If it appears safe, the rule is added to the active set with a confidence rating.

Formal verification is where the learned rules are checked against invariants. The Dispatcher's rules are not merely patterns observed from human behavior. They are formulas in a logic that the system can reason about. A rule that would enable path traversal is not discarded because it was observed to be safe in prior instances - it is discarded because it violates the path-confinement invariant by construction.

The Dispatcher becomes, over time, not a guard that blocks bad actions but a reasoning system that understands why actions are good or bad. Early versions learn from human decisions. Later versions learn from their own logical analysis. The human's role transitions from approver to auditor to, eventually, unnecessary oversight.

This is the bootstrap. The system begins dependent on human judgment because it has no basis for judgment of its own. Through accumulated decisions, it constructs a model of what is permitted and why. That model is the foundation for the deterministic symbolic engine that in v3.0.0 takes over the reasoning that the Dispatcher learned to perform.

The REPL as Cognitive Substrate

A REPL - Read, Eval, Print, Loop - is an interactive programming environment that reads an expression, evaluates it, prints the result, and loops back to read the next expression. It is the opposite of batch processing: where batch compiles and runs a program in one shot, a REPL works one expression at a time, with each evaluation building on all previous ones. The programmer defines a function, calls it, inspects the result, modifies it, and calls it again. The state accumulates. The session is the program.

In Lisp, the REPL is not a debugging tool bolted onto the language - it is the natural mode of interaction. The running image is the environment. When you evaluate (+ 2 2), the result 4 is printed, and you remain in the same image where + is defined, where previous definitions persist, where the next expression can reference anything that came before. There is no separation between development and execution. The REPL is not a simulation of the program - it is the program running.

Passepartout uses the REPL in this spirit, but elevated: it is not merely a tool for writing code, it is the mechanism by which the agent interacts with its own cognition - a loop that mirrors the perceive-reason-act metabolic cycle at the implementation level.

In the agent's cognitive architecture, the REPL serves three functions that are difficult or impossible to achieve through batch processing or stateless API calls.

First, the REPL enables verification before commitment. When the agent generates code, it does not write and forget - it evaluates in a running image, observes the result, iterates if incorrect. The feedback loop is tight: the time between writing and seeing the error is measured in milliseconds, not in the round-trip to a language server or a batch compiler. This is the "verification over hallucination" principle from the RLM paper made concrete: the agent tests what it writes before claiming it works.

Second, the REPL enables stateful exploration. The agent can define a variable, inspect it, modify it, redefine it. The exploration accumulates state across interactions. This is not a debugging session - it is the agent thinking with its hands, working through a problem by trying variations and observing outcomes, keeping the successful ones and discarding the failures.

Third, the REPL is a shared substrate. When the agent evaluates code, that code runs in the same image as the agent's own cognition. There is no process boundary between the agent and its tools. The REPL is not a subprocess the agent controls - it is a direct interface to the agent's own nervous system.

This is why the REPL becomes more important as the system matures. In early versions, it is a development tool. In v0.6.0 and beyond, it becomes a cognitive tool: the agent explores hypotheses by evaluating them, verifies the output of sub-agents by inspecting live state, and tests modifications before committing them to the knowledge graph.

Literate Programming as Discipline

The decision to use Org-mode as the source of truth for code, not just documentation, is not a ceremonial preference. It is a constraint mechanism that enforces better engineering habits at the cost of convenience.

The traditional development workflow is: write code, write comments, commit. The literate programming workflow is: write prose, write code, commit the Org. The order matters. The prose must come first not because of style guidelines but because the act of explaining what a function does before writing it forces clarity of thought that editing code directly does not.

When you must write a paragraph describing what a function does before you write the function, you discover the cases you have not considered. You find the edge conditions that are ambiguous. You realize that the function's name does not match its behavior, or that its behavior does not match your intent. The friction is not a bug - it is the mechanism by which thinking is enforced.

The one-function-per-block rule enforces granularity. A function that cannot be explained in a paragraph is a function that is doing too much. The block boundary is not aesthetic - it is architectural. It prevents the drift toward monolithic functions that accumulate responsibilities over time and become untestable, unmaintainable, and incomprehensible.

The tangle step enforces source-of-truth discipline. The .lisp file is generated from the Org file. This means the Org file cannot drift from the implementation. If the implementation changes, the Org must be updated to match. If the Org describes behavior that the implementation does not perform, the tangle produces code that does not match the Org description. Either way, inconsistency is visible and recoverable.

The evaluation gate enforces correctness. Every block can be evaluated independently in a running Lisp image. This means syntax errors are caught at authorship time, not at integration time. The function that compiles in isolation but fails in context is the function whose context dependencies were never made explicit. The evaluation gate forces those dependencies to surface.

Together, these constraints create a development experience that is slower in the small and faster in the large. Writing a new function takes longer because you must explain it. But debugging, maintaining, and extending the codebase is faster because every function has a human-readable explanation of its intent, every function is testable in isolation, and every function's source is always synchronized with its documentation.

The literate programming discipline is not about producing documentation. It is about producing code whose correctness has been verified by the act of explaining it.

The Evaluation Harness

SOTA parity is meaningless without measurement. A system that claims to match commercial agents must demonstrate it through reproducible benchmarks, not through feature checklists. The evaluation harness is the apparatus by which Passepartout proves its capabilities.

The industry standard for coding agents is SWE-bench: a corpus of GitHub issues paired with pull requests. The agent is given an issue, must understand the codebase, write a fix, and submit. Success is measured by whether the submitted PR passes the existing test suite. This tests the full chain: understanding, planning, code generation, verification, and multi-step reasoning.

Passepartout implements a native Lisp harness for this. A background thread clones repositories, feeds issues into the cognitive loop, tracks the resolution trajectory as an Org-mode headline tree, and scores success by test outcomes. The trajectory is persisted: when a resolution fails, the system can inspect where in the chain the reasoning broke down. The headline tree records the agent's thoughts at each step, making the failure auditable and the debugging human-assisted.

Beyond SWE-bench, the harness includes chaos testing. The system is subjected to resource starvation, concurrent load, and adversarial input. The deterministic engine must maintain safety invariants under pressure. The symbolic verifier must not deadlock or livelock. The probabilistic engine must degrade gracefully - if tokens are limited, it must still produce valid proposals that the deterministic engine can evaluate. Failure under chaos is a design flaw, not a benchmark anomaly.

The harness also supports regression testing on the skill set. Every skill is tested against a suite of known inputs and expected outputs. When a modification is proposed to any skill - whether through manual editing or the agent's own self-modification - the test suite runs first. A skill that fails its tests is rejected before it can propagate to the running image. This is not a convenience - it is the mechanism by which self-modification remains safe. The agent can propose changes, but the harness verifies them before the changes take effect.

Observability and the Thought Trace

When a human asks why the system made a decision, the answer must be findable. In most AI systems, the reasoning is ephemeral - it exists in the model's activations and disappears when the session ends. In Passepartout, every significant cognitive event is written to an Org buffer as it happens.

The thought trace is the agent's journal, written in parallel with its reasoning. When the probabilistic engine generates a proposal, the trace records the input, the prompt, and the raw output. When the deterministic engine evaluates it, the trace records which rules were checked, which passed, which failed, and why. When an action is executed, the trace records the timestamp, the user who approved it (if human-in-the-loop), and the outcome.

This is not logging in the traditional sense. Logs are forensically useful but are written in a machine format optimized for storage, not for human reading. The thought trace is written in Org-mode: headlines for major events, property drawers for structured data, tags for categorization. The human can open the trace in Emacs and navigate it like any other Org file. They can search for a specific decision, filter by time range, find all actions blocked by a specific rule, or see the complete trajectory of a multi-step task.

The trace becomes the foundation for the Dispatcher's learning. Every blocked action is in the trace. Every approved exception is in the trace. The human-in-the-loop decisions are in the trace. The system does not need to reconstruct what happened - it reads what happened from the trace it wrote.

Without observability, the system is a black box that happens to produce correct outputs sometimes. With observability, the system is auditable. The human can see why a decision was made, identify where the reasoning failed, and course-correct the system or its own behavior accordingly.

The MCP Strategy

The Model Context Protocol (MCP) is a standard for connecting AI systems to external tools and data sources. It defines how a client requests tools from a server, how the server exposes its capabilities, and how the client invokes them. The ecosystem is growing: MCP servers exist for GitHub, Slack, Postgres, filesystem access, and much more.

Passepartout connects to this ecosystem, but not by becoming a Node.js runtime. The architecture is: external MCP servers communicate via stdio or SSE to a Lisp-native MCP client that runs in the same image as the agent. The client is pure Common Lisp - it parses the JSON-RPC messages, invokes the tools, and presents results to the agent as Lisp data structures. There is no serialization overhead between the agent and the MCP layer, no process boundary, no impedance mismatch.

When the agent calls a tool via MCP, it receives a plist with the tool name, arguments, and result. The result is immediately usable by the agent's symbolic engine. When the agent generates a file, it can be written to the filesystem through an MCP filesystem server. When the agent needs to send a message, it can use an MCP Slack server. The agent does not need to know that these are MCP interactions - it sees only the plists that flow through its cognitive architecture.

The alternative is to build MCP wrappers in Python or TypeScript and bridge to Lisp via subprocess. This is what OpenClaw does: a Node.js runtime that manages MCP servers, with a bridge to the Lisp process. The bridge introduces latency, serialization costs, and a maintenance burden. The Node.js process must be kept running. The bridge must be maintained across Lisp and JavaScript runtimes. The cognitive architecture must handle errors that cross the process boundary.

Passepartout's native client is smaller, faster, and more maintainable. The MCP client is a skill, not a core component. It can be reloaded, replaced, or removed without restarting the agent. The agent can add new MCP tool integrations by loading new skills, not by deploying new infrastructure.

Local-First Architecture

Passepartout is designed to run on the user's machine, on their hardware, with their data, without requiring an internet connection. This is not a deployment option - it is an architectural commitment. The system must be able to reason, plan, and act using only the resources available locally.

The motivation is not merely philosophical. Cloud-based AI agents are economically incentivized to collect data, to train on user interactions, and to build lock-in through proprietary formats and network effects. When the agent runs locally, the user owns the hardware, owns the data, and can terminate the process without asking permission. There is no vendor that can change terms, no service that can go offline, no model that can be updated without consent.

Technically, local-first means several things. The LLM must be able to run on local hardware. Passepartout supports Ollama as a provider, which runs quantized models on CPU and GPU without requiring an external API. The vector database must be local. Passepartout uses its own org-object store, which is a folder of Org files that the agent already owns. There is no ChromaDB or Qdrant to install, no cloud vector service to authenticate with.

The symbolic engine does not require a network connection. The Prolog/Datalog reasoner that in v3.0.0 verifies neural proposals runs entirely in the Lisp image. The Dispatcher's rule synthesis does not call an external service. The agent can operate in a disconnected environment indefinitely, resuming full capability when connectivity is restored.

This does not mean Passepartout refuses to use cloud services when available and appropriate. It means cloud services are optional enhancements, not architectural requirements. The core is local. The user can choose to add cloud LLM providers for more capable inference, but the system functions without them.

Token Economics and Performance Advantage

This section analyzes how Passepartout's architectural decisions translate into token usage, latency, and cost versus competing agent designs (OpenClaw, Hermes, Claude Code).

The Core Insight: LLM as Expensive Resource, Not Default Engine

Passepartout treats the LLM as a resource to be minimized. Every operation is designed to reduce LLM dependency. Competitors treat the LLM as the core engine through which all operations flow. This is not a difference of degree but of architecture.

The three structural multipliers are:

Sparse tree retrieval — loading relevant subtrees (200-800 tokens per file) rather than full files (1,500-5,000 tokens) = ~5-10x reduction per file access

1. Deterministic safety — 9-vector dispatcher gate runs in pure Lisp (0 LLM tokens per verification) versus prompt-based guardrails (200-500 tokens per action) = infinite multiplier
2. REPL verification — catches errors in-image (milliseconds, 0 LLM tokens) versus LLM correction round-trips (500-2,000 tokens per retry)

These compound. A coding session touching 20 files, performing 10 actions, and triggering 3 errors saves ~50,000-100,000 tokens compared to the same session with Claude Code.

Per-Task Type Analysis

Coding (debugging, refactoring, PR review)

Operation	Passepartout	Claude Code	Hermes (3-agent)	Savings vs Claude
File access (30 files)	30 × 400 tok = 12,000	30 × 3,000 tok = 90,000	30 × 3,000 tok × 3 = 270,000	78,000 tok
Reasoning rounds (20)	20 × 3,000 tok = 60,000	20 × 4,000 tok = 80,000	20 × 3,000 tok × 3 = 180,000	20,000 tok
Error correction (5 caught by REPL)	0 (REPL)	5 × 1,000 tok = 5,000	5 × 1,000 tok × 3 = 15,000	5,000 tok
Safety verification	0 (deterministic)	500 tok/round × 20 = 10,000	200 tok/round × agents	10,000 tok
Agent coordination	0	0	3,000-5,000 tok/task	0
Total	~72,000 tok	~185,000 tok	~475,000 tok	~113,000 tok (2.6x)

Over a month of daily coding (20 sessions): ~2.3 million tokens saved. At typical API pricing ($2-15/M tokens), this saves $5-35/month.

Knowledge Management (Zettelkasten, research, note-taking)

Passepartout's strongest domain. The Org-mode native format and sparse tree retrieval create a 10-40x advantage because knowledge bases are the worst case for "load everything" architectures.

Operation	Passepartout	Competitor	Savings
Context assembly (500-node KB)	Peripheral outline + ~5 foveal nodes = 2,000-4,000 tok	Full serialization = 80,000-150,000 tok	40-75x
Semantic search (10 queries)	Vector lookup in-image = 0 LLM tok	LLM-assisted search = 5,000 tok	5,000 tok
Note creation (10 notes)	Deterministic Org writes = 0 LLM tok	10 × 800 tok = 8,000	8,000 tok
Total per session	~7,000 tok	~95,000-165,000 tok	~13-24x

Day-to-Day Life Management (calendar, tasks, reminders)

Operation	Passepartout	Competitor	Savings
Background maintenance	Deterministic heartbeat-driven = 0 LLM tok	Scheduled LLM calls or skipped	Variable
User interactions (30/day)	30 × 2,000 tok = 60,000	30 × 4,000 tok = 120,000	60,000 tok
Context queries by TODO/tag	Hash table scan = 0 LLM tok	LLM-based search = 2,500 tok	2,500 tok
Total per day	~60,000 tok	~122,500 tok	~2x

The defining advantage: background maintenance (compaction, archiving, link repair) costs zero LLM tokens. Competing systems either skip this or pay LLM costs for it.

Chatting (casual conversation)

Chatting is inherently LLM-bound. Passepartout's edge is privacy filtering before content reaches the LLM and slightly smaller context footprint. Token savings are marginal (~1.3x).

The Dispatcher Learning Curve: Cost Decreases Over Time

A unique architectural property: Passepartout's cost curve descends while competitors' ascends.

Passepartout: As the dispatcher accumulates deterministic rules from Human-in-the-Loop decisions, fewer actions require LLM proposals. A file write that initially triggered a full LLM proposal → dispatcher review → HITL approval → rule extraction loop eventually becomes a deterministic rule check. Each hardened rule permanently reduces future token costs.

Competitors: As context histories grow, safety instructions accumulate, and guardrails become more elaborate, each interaction costs more than the last. The only way to reduce cost is to cap context — sacrificing capability.

After 12 months of learning, Passepartout's core reasoning costs could drop to 40-60% of baseline, while competitors' costs rise to 125-140% of baseline.

The crossover point where Passepartout becomes structurally cheaper is estimated at 3-6 months depending on usage volume and task diversity.

Local LLM Viability

Reduced context requirements change which model sizes deliver acceptable performance:

Model	Passepartout Viability	Competitor Viability
Phi-3-mini 3.8B (4K ctx)	Viable for structured tasks	Context starvation
Llama 3.1 8B (8K ctx)	Comfortable daily driver	Marginal
Qwen 2.5 7B (4K ctx)	Viable for most tasks	Not viable
Mistral 7B (8K ctx)	Comfortable	Marginal
Llama 3.1 70B (128K ctx)	Overkill (but works)	Comfortable

KV cache memory scales with context length:

Context Window	KV Cache (Llama 3.1 8B, FP16)
4K tokens	~67 MB
32K tokens	~540 MB
128K tokens	~2.1 GB

Passepartout at 4K effective context: ~67 MB KV cache. Competitor at 128K: ~2.1 GB. A 7-8B model on an RTX 3060 Ti (8 GB VRAM) or MacBook (16 GB unified memory) is a practical daily driver with Passepartout. Competitors at full context require 16-32 GB VRAM or cloud APIs.

Open Questions and Risks

1. Retrieval accuracy is the bottleneck. If sparse tree retrieval loads the wrong subtree (low-similarity but causally relevant), the LLM makes unfixable errors. The architecture assumes embedding quality is "good enough" — this is untested at scale.
2. System prompt overhead can consume savings. Every think cycle iterates all registered skills and calls every system-prompt-augment function. With 20+ skills, a trivial interaction could carry 3,000-8,000 tokens of overhead before user input is even processed. This overhead is flat per-call, so it disproportionately affects short interactions.
3. Model size vs context quality. A 3.8B model with perfect context cannot match a 70B model on complex multi-file refactors regardless of context quality. Model size independently determines reasoning depth. The minimum viable model is likely 7-13B parameters for engineering work.
4. The 3-retry dispatcher loop. When the dispatcher rejects a proposal, the rejection trace feeds back to the LLM for self-correction (up to 3 retries). If the dispatcher rejects 30% of proposals, the effective token multiplier is 1.39x per action. At 50% rejection (plausible during early use), it is 1.75x. This penalty decreases as the dispatcher accumulates rules.
5. Competitor evolution. Sparse retrieval is not patentable. Claude Code, Copilot, and others will implement similar mechanisms. The architectural advantage is real but finite in duration. The deterministic safety gate is the harder-to-replicate differentiator.

Comparison Summary

Metric	Passepartout	Claude Code	Hermes	OpenClaw
Active context (tokens)	2,000-4,000	10,000-50,000+	5,000-15,000/agent	10,000-40,000
File access cost (per file)	200-800 tok	1,500-5,000 tok	1,500-5,000 tok × agents	1,500-5,000 tok
Safety verification cost	0 (deterministic)	200-500 tok/action	200-500 tok/action × agents	100-300 tok/action
Agent coordination cost	0	0	1,000-3,000 tok/task	500-2,000 tok/task
Error recovery cost	0 (REPL)	500-2,000 tok/retry	500-2,000 tok/retry × agents	500-2,000 tok/retry
Long-term cost trend	Decreasing	Increasing	Increasing	Flat/Increasing
Min viable local model	3-4B params, 4K ctx	30-70B params, 32K+ ctx	30-70B params, 32K+ ctx	7-13B params, 8K+ ctx
Min VRAM for local	4-6 GB	16-32 GB	24-48 GB	8-16 GB

Conclusion: Passepartout's architecture is designed to produce 2-3x token savings for coding, 13-24x for knowledge management, and 2x for life management at v1.0.0 maturity. The three structural advantages — sparse trees, deterministic safety, and REPL verification — compound. The critical risk is implementation gap: achieving the retrieval precision, dispatcher learning, and REPL integration depth required to realize the design.