- Who it helps
- Developers designing assistants that can recover from incomplete work.
- Availability
- The linked guide includes a Claude Code skill and standalone local examples. Model-connected behavior requires separate implementation and evaluation.
Separate controls for different failure modes
A new exercise tests approval tied to an exact revision and stops on exhausted attempts, elapsed time or a repeated result. It complements the existing citation and receipt exercises.
Claude Code extension model ↗Claude agent engineering: loops and graphs ↗
Run the stopping examples
Run the supplied examples, deliberately break an approval revision, and verify that the workflow holds the job for review.
What you can do
- Use the extension overview to distinguish reusable instructions from isolated workers, then run the guide’s local examples. Open reference ↗
Where the evidence stops
Only original deterministic helper programs were tested. No live multi-agent workflow or third-party runner was executed.
Sources read. Public evidence checked October 18, 2026. The product, external code and reported outcomes were not independently executed.
Practical guide
Skills, subagents, loops, graphs: choose the mechanism your task needs →
A mechanism decision table, reusable skill, evidence checks and runnable examples for bounded repair, missing receipts and approval-aware stopping.
Sources & further reading
Web · official
Claude Code extension model ↗
Vendor definitions. The architecture exercise and decision criteria are original instructional material.
Research authors / Anthropic · page checked 2026-10-18
Paper · paper
EvalResearchBench abstract ↗
Motivation for separating evaluator construction from held-out assessment; the receipt checker is original.
Research authors / Anthropic · page checked 2026-10-18
GitHub · implementation
Epoptes README ↗
Read the runner’s control commands, cycle limits and Codex budget limitations; no installation was performed.
ramigb · page checked 2026-10-18
Paper · paper
BizBench abstract ↗
Read the distinction between end-goal success and valid execution under business constraints.
BizBench authors · page checked 2026-10-18 · Published 2026-10-14
More sources reviewed (3)
Paper · paper
Assuring AI-written software: case-study abstract ↗
Motivation for checking omissions and weak proxies; no general reliability claim inferred.
Research authors / Anthropic · page checked 2026-10-18
Web · official
Claude Code skill locations and invocation ↗
Project SKILL.md location and slash invocation checked; the account-brief skill below is original.
Research authors / Anthropic · page checked 2026-10-18
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
Reusable instructions, isolated workers, bounded retries and explicit states solve different workflow problems.