M4 — ACE evolving playbooks

Paper: Zhang, Hu, Upasani, … Thakker, Zou, Olukotun — Agentic Context Engineering (ICLR 2026)

Linear: AGE-9 · Crate: chatty-playbook · Promotion: pending

Mechanism

Three roles over one playbook artifact:

Generator — runs task, reports which bullets were used (bullet_ids)
Reflector — tags bullets helpful/harmful/neutral from execution feedback (≤5 rounds)
Curator — emits JSON delta ops (ADD, UPDATE); never rewrites whole playbook

Merge is deterministic code — no LLM in apply(). Bullet format: [ctx-00263] helpful=1 harmful=0 :: <content> in named sections.

Grow-and-refine bounds size via de-duplication; paper uses loose thresholds (10K–100K token prune trigger) — in Chatty this becomes a production eviction policy, not a knob.

Two failure modes to reproduce as negative controls:

Brevity bias — optimizers drop domain heuristics (ACE cites GEPA’s shorter prompts)
Context collapse — monolithic rewrite shrinks artifact (AppWorld: 18K tokens → 122)

Chatty mapping

Paper concept	Chatty location	Status
Facts / procedures store	`memory_service`, `remember_tool`, `save_skill_tool`	Ships
Sectioned playbook	`chatty-playbook::Playbook`	To build
Delta merge	`chatty-playbook::apply`	Reserved
Grow-and-refine	`chatty-playbook::grow_and_refine`	Reserved
Generator	Existing running agent	Ships
Reflector / Curator	New `rig_agent::Agent`s	To build
Stable section order	`BTreeMap` serialization	Required for prompt-cache cost story

Product prize: playbook grows from execution feedback, not only when the agent calls remember.

Eval protocol

Stage A:

apply pure: ADD, UPDATE counters, no-op on unknown id, parallel deltas commute
Byte-stable serialization across no-op update (prompt-cache invariant)
Context-collapse control reproduces collapse curve
Label-free modes including degradation case (FiNER 70.7 → 67.3 without labels)

Stage B: AppWorld (primary), Finance (FiNER, Formula) — the human reviewer posts numbers.

ACE vs GEPA rollout comparison → AGE-21, not this module alone.

Production landing

Mechanism	Likely promotion
Deterministic `apply` + runtime playbook	Setting (scope TBD)
Playbook scope	Global / per-model / per-conversation (AGE-47)
Eviction policy	Default bounds required for desktop app (not paper’s 10K–100K knob)
Offline multi-epoch warmup	Optimizer feature, not request path

Reserved symbols

apply, grow_and_refine — see RESERVED.md.

Depends on

M0 Trace, M1 ReAct
Independent of M2 — can parallel with AFlow

Chatty Documentation