Skip to main content

Clarity Harness · Open Source · Launching Soon

Compound your context.

Every agent run makes the next one better. Incubate your agents — verdicts, not rewrites — until they earn autopilot, and your taste compounds right in the flow of work.

Coming soon · Free & open source

01 / 05 · RUN CAPTURE

Every run in one queue, review state attached.

The harness captures every agent run — story drafts, compliance checks, recon reports — grouped by agent with its review state on the row.

Scroll to explore · illustration of permitted learning

Old world · 01

You bought agents. You got slop.

  • Drafts pile up, unreviewed and unowned
  • Your best people babysit every artifact
  • “Done” quietly means “probably”

New world · 01

Every run in one queue, review state attached.

The harness captures every agent run — story drafts, compliance checks, recon reports — grouped by agent with its review state on the row. Nothing ships as a silent maybe.

Capture the run

The harness runs queue — agent runs for a payments platform with pending-review states
The runs queue — every agent, every run, every review state.

Old world · 02

You’re the eval. Forever.

  • Seniors read everything, line by line
  • Or rubber-stamp at 5pm under deadline
  • Judgment is the bottleneck — it doesn’t scale

New world · 02

Verdicts in minutes, receipts attached.

Approve, accept with edit, reject, defer — one keystroke each. The chain, the diff, and the source sit one tab away, so judging a run takes minutes, not meetings.

Bank reviewed judgment

Try it — judge a real run

MER-508 Chargeback evidence packet — assembled from the dispute timeline

Old world · 03

Your corrections evaporate.

  • The same mistake, caught every week
  • Fixes die in Slack threads
  • Agents never remember being wrong

New world · 03

Corrections carried forward. Forever.

Every correction is a reviewable diff — title, summary, acceptance criteria — banked into a versioned dataset that feeds the next regeneration. Fix it once and the whole harness learns.

Carry the correction forward

The harness delta view — an artifact version diff with additions highlighted, ready to approve or annotate
The delta — every version, every change, one diff away.

Old world · 04

You review forever.

  • Trust never accumulates
  • Week 12 reviews look exactly like week 1
  • Autonomy stays a demo

New world · 04

Agents level up — all the way to autopilot.

Advisor → Copilot → Autopilot. Promotion needs the full checklist, not a headcount: gates cleared with annotated traces, golden sets, and review coverage — the harness tracks the XP.

Review promotion gates

Agent standing page — Advisor, Copilot, and Autopilot tiers with promotion gates, XP, and the error-analysis quest log
Agent standing — tiers, promotion gates, and the quest log to autopilot.

Old world · 05

Your prompts rot in a doc.

  • The “latest” system prompt lives in five places
  • Nobody knows what changed, or why
  • Improvements can’t be rolled back — or repeated

New world · 05

The harness itself is versioned.

System prompt, skills, acceptance-criteria templates — every regeneration ships as a new harness version with a reviewable diff. v0.1 at 21 traces, v0.3 at 87: you can see exactly how your agent got good.

Compare harness versions

The PM Story Writer harness page — versions over time with trace counts, and the live system prompt
The Harness tab — versions over time, the live system prompt, the grading rubric.

Speed without compounding is just expensive slop.

The promise

Every verdict compounds.

The harness closes the loop: capture the run, deploy the agent, review with a verdict, learn from the correction. Today's annotation is tomorrow's default. Week one feels like review. Week six feels like leverage.

  1. CaptureEvery run recorded, receipts attached
  2. DeployThe agent runs your workflow
  3. ReviewYou give the verdict in minutes
  4. LearnCorrections banked into the next run

The promise

Every verdict compounds.

The harness closes the loop: capture the run, deploy the agent, review with a verdict, learn from the correction. Today's annotation is tomorrow's default. Week one feels like review. Week six feels like leverage.

Clarity Loop

Capture

Every run recorded, receipts attached

Deploy

The agent runs your workflow

Review

You give the verdict in minutes

Learn

Corrections banked into the next run

Continuously improving

Proof

We've helped real businesses with AI — from YC startups to $2B companies.

We can help you too.

The Harness direction · coming soon

Clarity Harness. Free OSS, coming soon.

harness_ui

The full harness UI

Mission control, run views, judgment queue — the exact interface in the screenshots above.

run_capture

Runs with receipts

Every run recorded: source input, agent chain, tool calls, decisions taken and not taken.

verdicts

The verdict system

Approve, edit, reject, defer — annotations persist as structured data, not Slack threads.

datasets

Versioned datasets

Corrections become a dataset with lineage — dev, staging, prod — that regenerates your agents better.

mcp_serve

MCP serve

Point Claude Code or any coding agent at your harness — orchestrator calls the agent, agent calls the skill.

setup_guide

Our setup playbook

The internal guide we used to stand up our own harness. Release scope and timing are not yet announced.

Free OSS Clarity Harness is coming soon. AI-native services start at $100k+. Clarity API enterprise pilots for company brain and hyper-personalization are custom quoted. Explore company brain.

Work with Clarity

Choose the work you need.

Operate your own tooling, engage a delivery team, or scope a context integration. These are distinct paths, not subscription tiers.

Built in the open

A community for builders who believe agents deserve better than vibes.

The harness ships free and open source because the thesis is bigger than a tool. This is the Clarity bet — and we're building a community of AI builders who share it:

◆

Subjectivity is the missing primitive

Autonomy doesn’t come from bigger models. It comes from knowing whose goals, whose judgment, and whose context a run answers to.

◆

Alignment is layered

Individual, team, and organization goals aren’t the same thing — a real harness reconciles all three instead of pretending one prompt covers it.

◆

Closed loops beat bigger prompts

True autonomous agents are grown, not prompted: run → verdict → learn → run again, with the human judgment banked every cycle.

If that's your thesis too, don't just star the repo — come build the loop with us.

Scaling past your own harness? Company-wide harnesses are the layer we build with design partners — see the company brain →

Free OSS Harness · coming soon

Join the Harness waitlist.

Clarity Harness is the free, open-source builder path, coming soon. Join the waitlist for release updates.

Coming soon · free & open source