Skip to content

README.md

Latest commit

 

History

History
89 lines (67 loc) · 9.31 KB

README.md

File metadata and controls

89 lines (67 loc) · 9.31 KB
Raw
topic agents
type market-research
status research-complete
last-validated 2026-07-29
superseded-by
related-docs 928, 2036, 994
original-query https://x.com/zodchiii/status/2082408590460006732 - Anthropic engineer: build a system with Opus 5 that prompts itself, not prompt Claude
tier STANDARD

2127 - Loop / Harness Engineering (Anthropic "systems that prompt themselves") - ZAO alignment audit

Goal: Take the viral "you're not supposed to prompt Claude, you build a system that prompts itself" talk and answer the only question that matters for us: what does ZAO already do, and what are the 2-3 concrete tune-ups worth making.

Key Decisions (recommendations first)

# Decision Why Owner
1 Do NOT build anything new from this - ZAO already practices loop/harness engineering. Doc 928 + .claude/rules/agent-loops.md (36 rules) + the live ZOE loops (orchestrator-tick.ts, work-loop.ts) ARE the harness the talk describes. Treat the talk as external validation, not a to-do. The talk's whole thesis (self-prompting observe-plan-act-reflect loops, memory files, stop conditions, human gates) is already the ZAO operating model. Building "loop engineering" would duplicate what ships. @Zaal
2 Audit the "claude.md tax." The talk's one genuinely actionable metric: poorly-structured context files cost ~14% of productivity (they cite up to 60% lost to bad context). ZAO's CLAUDE.md + MEMORY.md + 10 .claude/rules/*.md are LARGE and auto-load every session. Measure whether they help or tax. Doc 2036 already flagged MEMORY.md size + prompt-cache as the biggest context lever. This talk quantifies the cost and makes the audit concrete. @Zaal
3 Adopt the 3-level maturity ladder as a shared vocabulary for ZAO loops, nothing more. L1 observe-act (minutes), L2 task-decomposition (a day), L3 autonomous multi-day with checkpoints + human gates (a week). Map each ZAO loop to a level so their scope is legible. Cheap, no code, closes the "which loops are running and what can they do" gap (workflow-discipline rule 2). @Zaal
4 Note where ZAO is AHEAD, not behind: persistent cross-run learning. The blog explicitly says the talk does NOT detail agents learning across separate runs. ZAO's rule 6/9 ("persist lessons to the repo, not just memory") + the 36 fold-back rules in agent-loops.md ARE cross-run learning. Keep doing it. Don't cargo-cult a talk that is behind us on the one axis that compounds. @Zaal

The talk (what it actually says)

A clip (17 min in the zodchiii version; a 45-min version circulates attributed to a different speaker - attribution is muddy across the viral reposts, some say "he", some "she"; do not cite a name) of an Anthropic engineer arguing the 2026 skill is loop / harness engineering, not prompt engineering.

Core thesis, verbatim: "You're not supposed to prompt Claude. You're supposed to build a system that prompts itself."

The canonical loop:

  1. OBSERVE - read codebase, tests, error logs
  2. PLAN - decide the next action from observations
  3. ACT - make changes, run tests, commit
  4. REFLECT - evaluate, find gaps, adjust
  5. REPEAT - until done or timeout

Runs autonomously for hours/days with no human between cycles. Single prompts fail because they can't iterate, lose context, and bottleneck on the human.

Supporting pieces:

  • Memory files - architecture map, code-style rules, testing strategy, deploy process, business/security context. Bad ones = the "14% claude.md tax."
  • Task decomposition - break a feature into sub-tasks, each its own observe-plan-act cycle.
  • Stop conditions - max turns (20-50 typical), circuit breakers on repeated failure, token monitoring, human-approval gates for high-risk actions.
  • 3 implementation levels - L1 simple observe-act (~1hr), L2 multi-step decomposition (~1 day), L3 autonomous multi-day with checkpoints + human gates (~1 week).

Claimed numbers (Anthropic, by May 2026): 8x daily code output, 80%+ of merged production code authored by Claude, 76% success rate on open-ended software tasks, up to 60% of developer productivity lost to poorly-structured context.

Findings: the ZAO map (talk concept -> where it already lives)

Talk concept ZAO already has it Evidence
"System that prompts itself" The ZOE orchestrator loop + relay bridge built THIS session; the work-loop bot/src/zoe/orchestrator-tick.ts, bot/src/zoe/work-loop.ts
Observe-plan-act-reflect agent-loops.md rule 2 (read state before acting) + rule 1 (ground truth close) + rule 10 (learn online periodically) .claude/rules/agent-loops.md
Memory files CLAUDE.md + AGENTS.md + MEMORY.md + 10 .claude/rules/*.md, all auto-loaded repo root + .claude/rules/
Sub-agent split rule 7 (subagents for bounded research/isolation, inline for the hot path) agent-loops.md rule 7; doc 928 practice 8
Stop conditions / cost ceilings rule 5 (hard cost + iteration ceilings), daily caps, file-locks, empty-queue=zero-spend agent-loops.md rule 5; orchestrator-tick.ts DAILY_CAP + lock
Human gates for high-risk rule 8 (PR-only + human gate as circuit breaker); outbound/on-chain/spend gated agent-loops.md rule 8; workflow-discipline.md
Circuit breaker on repeated failure rule 30 (boot-verify hard-fail), rule 32 (verify the right commit), auto-rollback in zoe-autodeploy agent-loops.md rules 29-32
Persist lessons across runs rule 6/9 (persist to .claude/rules, not session memory) - talk does NOT do this; ZAO does 36 fold-back rules in agent-loops.md

Bottom line: of 8 concepts in the talk, ZAO already ships 8. Doc 928 (2026-06-30) is the ZAO distillation of Anthropic's earlier "Building Effective Agents" + "Effective Harnesses for Long-Running Agents" - this talk is the same lineage, restated for the Opus 5 era with a punchier name ("loop engineering") and marketing numbers. It is confirmation, not new information.

The genuine delta (what is actually worth doing)

  1. The claude.md tax is measurable and ZAO is exposed to it. ZAO's context stack is unusually large (a multi-thousand-line CLAUDE.md, a big MEMORY.md, 10 rule files, all loaded every session). The talk says bad context files cost ~14-60% of productivity. This is the one place the talk could save ZAO real cap. Pair with doc 2036 (context-hygiene) - measure, then trim what taxes.
  2. Name the levels. ZAO runs several loops (ZOE orchestrator, work-loop, autodeploy, research loops) at different autonomy. Tag each L1/L2/L3 so "what is running and what can it do" is legible (workflow-discipline rule 2). Zero code.
  3. Nothing else. No new loop, no new bot, no framework. The talk validates the road ZAO is on.

Also See

  • Doc 928 - the ZAO agent-loop rulebook (the earlier Anthropic guidance, already distilled into agent-loops.md)
  • Doc 994 - ZAO's own loop-engineering taxonomy (same thesis, already mapped)
  • Doc 2036 - context hygiene + cost discipline (MEMORY.md size, prompt-cache) - the claude.md-tax audit lives here

Next Actions

Action Owner Type By When
Audit the "claude.md tax": measure whether CLAUDE.md + MEMORY.md + rules help or bloat a session, trim the dead weight - land the trim as a PR @Zaal PR 2026-08-12
Tag each live ZAO loop (ZOE orchestrator, work-loop, autodeploy, research) with an L1/L2/L3 autonomy level in agent-loops.md @Zaal PR 2026-08-12
No new loop/bot from this doc - decision recorded; revisit only if a specific ZAO gap appears @Zaal Decision wontfix

Sources

  • zodchiii tweet (the source) - [FULL] - fetched via fxtwitter: text + engagement (35 favs, 2827 views, posted 2026-07-29). The embedded 17-min video is the primary artifact.
  • explainx.ai - Harness Engineering breakdown - [FULL] - full structured summary of the talk: observe-plan-act-reflect, memory files, stop conditions, the 8x/80%/76%/14% numbers, the 3-level ladder.
  • The embedded talk video (video.twimg.com/.../T8FIy-gf-IpHFRZg.mp4) - [FAILED] - zao-ingest.sh could not transcribe the direct twimg mp4 (it fell back to on-site-transcript extraction and produced nothing). The explainx blog is a full textual substitute for the talk's content, so the doc is not written off a snippet.
  • Viral reposts confirming the concept + name discrepancy: Vikas gupta, Anatoli Kopadze, Rahul - [PARTIAL] - titles/snippets read via WebSearch (not each thread's full tree); used only to confirm the concept is widely circulating and that speaker attribution is inconsistent across reposts.
  • ZAO grounding (code is ground truth): .claude/rules/agent-loops.md (36 rules), bot/src/zoe/orchestrator-tick.ts, bot/src/zoe/work-loop.ts, Doc 928 - [FULL] - read directly this session.