docs: add subagent iteration workflow to AGENTS.md

projects/AGENTS.md

@@ -1,5 +1,27 @@
 # AGENTS.md
 
+## Subagent iteration workflow
+
+When delegating work to subagents, follow this iteration loop:
+
+1. **Specify** — define the task clearly: what files to change, what the
+   output should look like, what constraints apply (read-only vs. build)
+2. **Receive** — review the subagent's output for correctness, edge cases,
+   style, and consistency with the existing codebase
+3. **Critique** — send back specific issues: "Block 3 has an off-by-one
+   error", "The README doesn't explain the --package flag", etc.
+4. **Iterate** — repeat steps 2-3 until zero issues remain. Do not accept
+   work that still has bugs, documentation gaps, or untested edge cases
+5. **Accept** — only when you cannot find any more issues
+
+Key principles:
+- The subagent implements, the architect reviews. Do not switch roles.
+- Every review pass must find fewer issues than the previous pass
+- If a subagent introduces new bugs while fixing reported ones, reject
+- Zero issues means zero: no style nits, no missing edge cases, no
+  unclear prose, no dead code
+- This applies to code, documentation, tests, and config files equally
+
 ## Development Cycle (every change)
 
 0. **Start the runtime** — boot the Lisp image that loads your project.