autoharness

autoharness

autoharness is a self-learning skill layer for Claude Code. It learns skills from your real sessions, merges same-scenario ones instead of stacking near-duplicates, updates them in use, and prunes any that stop getting used — so the layer stays clean on its own, touching only the skills it wrote itself.

Same model, different harness — 42% → 78% on CORE-Bench (HAL). The harness does much of the work (swyx's Big Model vs Big Harness), yet it's still rebuilt by hand every model generation. autoharness bets one slice of it — the skill layer — can maintain itself.

Learns from real work	Each episode is distilled into a skill from the session you were already having — no separate data-collection or replay loop.
Groups, doesn't just pile up	A new episode doesn't always add a skill — the reflector compares it against what's there and folds same-scenario skills into one, so the layer consolidates by category instead of accreting near-duplicates.
Validated in use, not on a benchmark	A skill survives by being adhered to in later turns (invocation rate), not a held-out score. No oracle on the active path, and no tokens spent on a dedicated eval.
Only its own skills	Touches only the skills it generated through this plugin — everything else, whether you wrote it or installed it, is left completely alone.
Evidence kept for later	Every create/update logs its scenario and decision to a per-skill ledger — the raw material to build a benchmark from real usage if you ever want one.

Install

Requires python3 on your PATH — autoharness runs entirely as Python (zero third-party dependencies); its hooks and MCP server won't fire without it.

/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharness

Run /reload-plugins (or restart Claude Code), then check it's live:

/plugin    # autoharness@autoharness — enabled
/mcp       # stage_skill — connected

Zero config. It now watches your sessions and lands learned skills into .claude/skills/ in the background. Cadence and lifecycle thresholds are tunable — see Configuration.

Update

/plugin marketplace update autoharness              # pull the latest release
/plugin update autoharness                          # update the installed copy
/reload-plugins                                     # apply in the running session

/reload-plugins re-arms the hooks and the MCP server in place; restarting Claude Code works too. The installed copy is cached by the version in plugin.json; releases always bump it, which is what makes the update take effect.

Uninstall

claude plugin uninstall autoharness@autoharness     # stops the hooks + MCP server
claude plugin marketplace remove autoharness        # optional — also drops the install source

Uninstalling only stops it from running — the skills it landed and its own state live outside the plugin and stay on disk. To clear those too, delete its state dir (~/.claude/autoharness/ global, <repo>/.claude/autoharness/ per project) and the self-authored skills under .claude/skills/ (each carries a self-authored ledger marker, so they're easy to tell from yours). Your own skills are never touched.

Configuration

Every knob is an AUTOHARNESS_* environment variable with a built-in default — nothing to configure unless you want to change the pace.

Variable	Default	What it does
AUTOHARNESS_REFLECT_EVERY_N	3	Reflection cadence: a background reflection run fires every N host turns. Lower = learns faster, spawns more child sessions.
AUTOHARNESS_DIGEST_EXCHANGES	20	How many exchanges before the episode window are compressed into the reflector's prior-context digest (text + tool names only).
AUTOHARNESS_MATURITY_PROJECT	100	Probation gate, project layer: a skill only graduates — starts counting against capacity and becoming evictable — after this many requests have arrived in its layer since it landed. Until then it's recalled as usual but can't be archived.
AUTOHARNESS_MATURITY_GLOBAL	300	Same gate for the global layer — higher because a global skill loads in every project.
AUTOHARNESS_CAPACITY_PROJECT	50	Cap on mature skills in the project layer. Capacity contention is the only death: nothing is archived until the mature pool exceeds this, then the lowest invocation rates go first.
AUTOHARNESS_CAPACITY_GLOBAL	20	Same cap for the global layer — smaller because its blast radius is every project.

Set them in the environment Claude Code launches with — either the shell (export AUTOHARNESS_REFLECT_EVERY_N=10) or the env map in .claude/settings.json:

{ "env": { "AUTOHARNESS_REFLECT_EVERY_N": "10" } }

Hooks read the environment on every event, so a change applies from the next session. The defaults are deliberate placeholders pending empirical calibration (tracked under experiments/); size caps on captured windows and staged skill bodies are fixed constants, not env knobs.

How it works

A learning pipeline runs beside the host and stays off its recall path — symbols are plain native skills, recalled by the host's own name-and-description mechanism as if a human had written them.

_{Diagram source: docs/assets/pipeline.mmd — re-render to pipeline.svg after editing.}

Component	Role
CAP · capture	Hook-driven dumb pipe: grabs each turn (user input, agent output, tool I/O), redacts at egress, points back at the host log instead of copying it.
REF · reflect	At an episode boundary, receives the current episode window in full detail (the last N turns, tool I/O included) plus a compressed digest of the exchanges before it (text and tool names only), reads the existing skill index, and decides add / merge / patch / drop a support file / delete — emits an intent (body, delta, or path, plus reason and evidence). Proposes only; no write tools.
promoter · validate·store	The only writer. Lints the intent in memory (safety, structure, ledger, completeness, self-authored-only) and on pass does an atomic rename into the live skill directory.
MNG · lifecycle	Daemon-free: recomputed lazily at session start, once per session. Ranks symbols by invocation rate — calls over the requests that arrived since the symbol was created, so the measure is opportunity-relative and a closed laptop doesn't age anyone out (the wall-clock replacement). New symbols sit in probation until they've had a fair sample of requests: recalled as usual, but neither counted against the cap nor evictable. Capacity contention is the only death — nothing is archived until a layer's mature pool exceeds its cap, then the lowest rates go first. Archives, never deletes: an archived symbol is a directory moved out of recall, and moving it back revives it.
LED · ledger	Per-symbol append-only sidecar: why each symbol was born or changed, with evidence and a reflection watermark. Kept out of the skill body so recall stays clean.

Walkthrough: watching it learn

Everything autoharness does lands on disk as plain files — a demo is just opening them in the right order. For a fast-paced run, speed up the loop first (see Configuration):

{ "env": { "AUTOHARNESS_REFLECT_EVERY_N": "1",
           "AUTOHARNESS_MATURITY_PROJECT": "5",
           "AUTOHARNESS_CAPACITY_PROJECT": "2" } }

1 · The pipeline running. Work a few normal turns on anything non-trivial (debug something, figure out a workflow). Every Nth turn a background reflection fires — nothing blocks your session. Its bookkeeping is visible in the state dir:

ls .claude/autoharness/        # per project — ~/.claude/autoharness/ for the global layer
  requests                     # layer request counter (MNG's denominator)
  session-<id>                 # per-session turn count toward the next reflection
  offset-<id>                  # byte watermark: where the last captured window ended
  intents/                     # queued skill proposals awaiting the promoter

2 · A skill is born. After a reflection lands, a new folder appears under .claude/skills/ (project) or ~/.claude/skills/ (global — for techniques that aren't repo-specific). Use ls -la: the interesting files are hidden.

.claude/skills/<name>/
  SKILL.md                     # the skill itself — plain native format, nothing proprietary
  .ledger.jsonl                # LED: why it was born / changed (append-only)
  .sidecar.json                # lifecycle counters MNG reads
  references/evidence-*.md     # the transcript slice that justified each ledger entry
  scripts/ templates/ ...      # optional support files the reflector attached

3 · LED — the paper trail. cat .ledger.jsonl — one JSON line per lifecycle event:

{"action": "create", "reason": "User asked about the correct command to update a plugin ...", "evidence": "references/evidence-21cd22cc.md"}
{"action": "patch",  "reason": "User discovered /reload-plugins is required in-session ...",  "evidence": "references/evidence-1a4ec51d.md"}

action + reason + evidence — and the evidence file is a real, redacted slice of the session that taught it, materialized by the promoter (content-addressed, so the model never names files). This is the "evidence kept for later" from the table above.

4 · An update, not a duplicate. Hit the same scenario again with a correction ("that's missing a step") and let the next reflection run. The skill layer does not grow a near-duplicate: the existing skill's SKILL.md changes and its ledger appends a patch/update line — the two-line ledger above is a real example. git diff on a project-layer skill shows the edit.

5 · Recall is the host's, untouched. Landed skills load like hand-written ones — same name-and-description recall, no autoharness code on that path. When one fires, calls in its .sidecar.json ticks up: that adherence count is the validation signal.

6 · Retirement is an archive, not a delete. Once a layer's mature pool exceeds capacity, the lowest-invocation-rate skills move — folder, ledger, evidence and all — to .claude/skills/.archive/<name>/, out of recall. Moving the folder back revives it, history intact. With the shrunk knobs above this fires within one session; at defaults it takes hundreds of turns.

7 · Yours are never touched. Every autoharness-authored skill carries the ledger marker; anything without it — skills you wrote or installed — is invisible to the promoter and MNG.

How it compares

A self-learning skill layer can be validated against a held-out benchmark, or against its own use. autoharness takes the second — cheaper, and it works on a live host doing open-ended work where no benchmark exists.

| | Grow unbounded | Offline-gated self-edit (Self-Harness) | Timer + daemon (hermes-agent) | autoharness | | --- | --- | --- | --- | --- | | Bounds the skill layer | No | Yes | Yes | Yes | | Validation signal | None | Held-out benchmark score | Wall-clock inactivity | Adherence in use | | Needs a benchmark / oracle | No | Yes | No | No | | Needs a resident daemon | No | No | Yes | No |

Acknowledgements

NousResearch/hermes-agent — studying its auto-skill-creation and memory-consolidation design helped sharpen autoharness's adherence-based, daemon-free take.

Built by Tigerless Labs.

License

MIT