TraceForge agentic trust

TraceForge.
Judge the outcome,
not the spend.

Did your agents’ spend buy finished, on-policy work? TraceForge tells you.

See How It Works Book A Demo
TraceForge reads
  • Agent-agnostic
  • Claude Code
  • OpenAI Codex CLI
  • GitHub Copilot CLI
TraceForge · Session Verdict sess_8a21f sess_c04d9 sess_1f77b
Refactor billing module tests · claude-code · 42 tool calls · $3.80 Migrate auth schema · codex-cli · 118 tool calls · $14.20 Fix flaky CI job · copilot-cli · 61 tool calls · $6.10

checking task completion… checking policy adherence… checking topic drift… pricing the outcome… verdict rendered

  • Task completion holds partial failed evaluating
  • Policy adherence holds holds failed evaluating
  • On-topic use holds drifted holds evaluating
  • Dollars-to-outcome holds partial failed evaluating
Verdict Complete Partial Failed Judging… Spend → outcome $3.80 → done $14.20 → 60% $6.10 → wasted —
The question

Did the agent
actually do the job?

  • authorized — the action was allowed to run
  • unknown — did it finish, on policy, on mission?

Authorization allows an action. It doesn’t confirm the work got done. TraceForge is the judge that steps in after the action.

How it works

How TraceForge works: ingest, structure, judge.

Nothing in the request path. No inference sold. TraceForge reads the logs after the fact and renders a verdict per session.

  1. 01

    Ingest

    Reads any agent’s session logs — nothing in the request path, no inference sold.

    source: claude-code · codex · copilot · custom
  2. 02

    Structure

    Links every model call, tool use and retry into one causal session.

    calls → tools → retries → outcome
  3. 03

    Judge

    Scores each session for completion, policy and topic — then prices the result.

    verdict: complete | partial | failed
What TraceForge judges

One verdict per session.

  • 01 Task completion Was the task done, partial or failed — not just what it cost.
  • 02 Policy adherence Did the agent stay inside its guardrails.
  • 03 On-topic use Did the session drift from the team’s actual work.
  • 04 Dollars-to-outcome How much spend bought failed or off-policy work.
TraceForge · Verdict Record rec_41c0e rec_9b2d7 rec_0e88a
Session Add rate-limit middleware Migrate user table Fix flaky CI job
  • 01 Task completion Middleware merged, tests green. 3 of 5 migration steps done. Job still flaky at end of session. pass partial fail judging
  • 02 Policy adherence Stayed inside its guardrails. Stayed inside its guardrails. Force-pushed to main — not allowed. pass pass fail judging
  • 03 On-topic use Every step served the task. Wandered into unrelated tests. Stayed on the CI problem. pass partial pass judging
  • 04 Dollars-to-outcome $4.10 bought finished work. $19.40 bought about half the job. $7.30 bought nothing usable. pass partial fail judging
Verdict Complete Partial Failed Judging… Spend $4.10 $19.40 $7.30 —
The loop-back

TraceForge verdicts become authority.

Two agents with identical identity earn different authority because their history is different. TraceForge’s verdicts feed TIM and the Trust Graph — performance shapes future permission.

ops-deploy · identical identity billing-refunds · identical identity ci-fixer · identical identity
agent://ops-deploy-01
Sessions judged complete
Authority granted
agent://ops-deploy-02
Sessions judged complete
Authority granted
agent://refunds-01
Sessions judged complete
Authority granted
agent://refunds-02
Sessions judged complete
Authority granted
agent://ci-fixer-01
Sessions judged complete
Authority granted
agent://ci-fixer-02
Sessions judged complete
Authority granted
Attempt TraceForge Outcome

With TraceForge, pay for outcomes,
not attempts.

Get TraceForge Book A Demo