APE started with nine named agents. Two months of building APE with APE collapsed the roster to four. Here's who's on the team, and the honest story of who didn't make it.
The unexamined life is not worth living.
Clarifies requirements through dialog. Where most agents guess, SOCRATES asks. Leaves behind a diagnosis of the problem, captured before a single line of code is written.
diagnosis.md
Divide each difficulty into as many parts as feasible and necessary to resolve it.
The Cartesian method — divide, order, verify, enumerate — applied to planning. Replaces two earlier apes (SUNZI and VITRUVIUS). One method does both their jobs, and it's testable.
plan.md
古池や
the old pond
a frog leaps in — plop.
Minimal, beautiful implementation under the constraints of the tests. Like a haiku: the constraint is the form itself. The craft happens inside the limits.
code + commits
Reads the cycle's artifacts and proposes mutations to APE itself — changes to the framework, the transition contract, the other apes. The only ape that improves APE. Off by default; when enabled, files its proposals as issues in this repo (evidence: #54).
Observation nº 54 — collapse of the roster from nine to four; three absorbed, one replaced, pattern consistent across six cycles.
repo issues
The original lore.md sketched nine apes. After two months of building APE with APE, the roster collapsed to four. This is honest accounting, not a roadmap to revive the deferred ones. Most were absorbed by simpler agents that turned out to do the job better.
| Lore agent | Status | What happened |
|---|---|---|
| SOCRATES | Active | ANALYZE — implemented as the mayéutica agent. |
| DESCARTES | Active | PLAN — replaces SUNZI's strategy and VITRUVIUS's WBS in one Cartesian method. |
| BASHŌ | Active | EXECUTE — replaces ADA's TDD with techne: functional beauty under constraints. |
| DARWIN | Active | EVOLUTION — implemented; produces concrete evidence (e.g. issue #54). |
| MARCOPOLO | Absorbed | Document ingestion handled by SOCRATES and the memory-read skill. |
| VITRUVIUS | Absorbed | WBS / decomposition is part of DESCARTES's plan phase. |
| SUNZI | Replaced | DESCARTES's method is more explicit and testable than strategic prose. |
| GATSBY | Absorbed | Test pseudocode lives in plan.md, written by DESCARTES. |
| ADA | Replaced | BASHŌ's techne replaces explicit TDD as a separate phase. |
| DIJKSTRA | Future skill | Quality-gate becomes a pre-pr-review skill inside END, not a separate agent. |
| BORGES | Future skill | Schema validation becomes ape memory validate, not a standalone ape. |
| HERMES | Materialized | State transitions are now ape state transition — a CLI command, not an agent. |
The framework wants fewer, sharper agents, not more.
Each absorption came from a real cycle where two agents were doing what one could do better. MARCOPOLO and the memory-read skill were doing the same reading SOCRATES already did during ANALYZE. VITRUVIUS's decomposition happened inside DESCARTES's method. GATSBY's test pseudocode lived naturally in plan.md.
This is antifragility applied to the framework itself: APE got better by running against itself and discarding what didn't earn its weight. Every deprecation is logged in a DARWIN evolution issue, with reasoning. Nothing disappears silently.