agent-harness GitHub
Menu

Overview

agent-harness Agent Harness Blueprint Blueprint plugin product brief The Harness Book Deny-floor known limitations Licensing scope and decision Replay tool product brief Agent Harness roadmap Harness Specs

Docs

Advisory agent-operations measurements Mapped authority-document drift probe Mapped authority drift: pinned local trial Benchmarks and measured evidence Corpus path boundary Advisory entropy measurements Learning evaluator evidence boundaries Outcome stream validation System comparisons and canary evidence Advisory logical-job accounting Replay lab Python check inventory Replay CLI input identity and timeout limits Replay input decoding failures Replay report output boundary Replay process policy-root aliases Replay supervisor marker scope Review evidence and acceptance oracles Offline reviewer-packet diagnostics System state Process-fixture startup is not containment evidence

Agentic

Architecture additives — perfect assistant → estate owners Perfect autonomous agentic assistant — estate synthesis

Evals

Harness extension notes Evaluation merge-boundary contract: v0 Merge-gate model Outcome ledger Review under flood Agentic development evaluation taxonomy: v0

Evals / Examples

Worked evidence packet: a private product PR

Extraction

Replay v0 extraction inventory

Freeze

Legacy deny-floor v1 freeze record

Maintenance

Managed Codex agent publication Sync input and Doctor session roots

Observability

Coding-agent loop observability v0 Offline source-record identity inspector

Observability / Adapters

Claude Code to agent-loop v0: mapping note

Review evidence

Review/oracle wave: source ledger

UX evaluation

Frozen observation verifier Evidence-bound issue seeds Advisory product UX evaluation Offline scenario binder Scenario authoring contract v0 Taskdeck-first adapter sketch Workspace handoff acceptance profile Workspace receipt lab

Review/oracle wave: source ledger

Checked 2026-09-23. This is a provenance ledger, not a new research source of truth or a replacement for the original brief. Links to private workflow sources remain private.

B0: requested Deep Research brief, unavailable

The commissioning request names review under PR flood, acceptance/oracle design, scarce-user validation, and agent-induced entropy. The accessible conversation contains the research request and a launch acknowledgement, but no completed brief, bibliography or durable artifact link. A targeted conversation/Library and personal-context search did not recover it.

Therefore no finding here is attributed to the completed brief. Its existence or contents are not disproved; its citation is unresolved. Before promoting these drafts as research-derived policy, attach the actual artifact/link in the PR discussion and reconcile its claims against this ledger. Do not invent a title, quotation, result or bibliographic entry. The source document should remain in its existing home when recovered.

Inspected estate facts

ID Source / snapshot Fact used Boundary
E1 agent-harness#101 and CLAUDE.md at 8bf4cda Consolidate review doctrine; floor/Doctor/replay responsibilities are distinct Issue history does not override current tier/laws
E2 BENCHMARKS at 8bf4cda Recorded results require bounded inputs, methods, results and limitations Historical entries were read, not rerun
E3 WORKFLOW_DESIGN at b41c86a Slice, review and evidence procedures already exist; canonical sources render into two runtime adapters Private source; source parity is not native activation
E4 Taskdeck dogfooding at 622d9d8 and #1947 Separate real-use database, existing checkpoint, observed friction and attribution problems Prior observations are not new pilot results
E5 agent-harness#281, PR #289, claude-config#308, agent-hq#14 UX creation, advisory UX auditing and their packaging have separate owners Not reimplemented in this wave
E6 agent-harness PR #288 and #233 Ops measurements and workstream-cap reconciliation already have homes Draft ops plans are not shipped behavior
E7 llm-release-gate README, inspected 2026-09-23 Existing offline fake-provider baseline/candidate gate labels exact versus heuristic metrics Demo success is not live-model or semantic acceptance proof; no integration added

Additional primary references, independently checked

These are supplemental references, NOT a recovered bibliography from B0.

ID Reference and date Supported statement Limit
R1 Google: Small CLs, undated guidance, accessed 2026-09-23 Prefer coherent, self-contained changes with related tests; size is conceptual, not just lines Practitioner guidance, not an experiment on this agent estate
R2 Google: Standard of Code Review, undated guidance, accessed 2026-09-23 Balance forward progress with improving code health rather than demanding perfection Does not set this estate’s review count, severity or merge authority
R3 Kleppmann et al.: Local-first software, April 2019 essay / Onward! 2019 paper Local ownership, privacy and user control are design goals Does not prescribe the telemetry retention period or dogfood thresholds below

New recommendations, not established findings

The compact reviewer packet, bounded control trials, drift calibration dataset, manual Taskdeck comparison and their suggested sample/time budgets are design proposals from this reconciliation. They have not been tested for adoption, reviewer-effort reduction, precision, product usefulness or long-run retention. No vendor performance claim is used. A good local pilot can justify the next bounded experiment, not universal policy or automated enforcement.