Tracing and evaluation identify failures. The Harness turns that evidence into scoped, validated rules agents can use to recover from them.
The Harness turns trace evidence and evaluation failures into structured repair work, connecting detection to improvement.
Learned rules live in a persistent workspace, scoped by task, workflow, or domain instead of disappearing with one session.
Every candidate is tested through developer-controlled task replay. Only rules that improve evaluated outcomes become trusted.
Validated rules stay available to relevant agents so known failures can be avoided instead of diagnosed and repaired again.
Scoped rules, restricted tools, validation gates, and human policy keep self-repair bounded and auditable.
Reuse scoped rules across the fleet to reduce manual repair on low-cost models and achieve higher success rates.
The Repair Harness evaluates traces, posts failure notices, writes scoped rules for the next turn, and validates what becomes trusted.
Trace evaluation
Each new trace is scored for task completion and coherence, while the trajectory is watched for stalls or regressions. Suspected failures trigger deeper tool and argument checks, with optional planning and efficiency diagnostics.
Workspace mailbox
A confirmed breach becomes a structured notice in the workspace mailbox, preserving the trace and metric evidence needed for repair.
Repair Harness
The Repair Harness groups related notices, inspects the flagged traces, and writes a scoped rule to the shared workspace. After settlement, the task agent can pull it on the next turn.
Evidence gate
Replay or live trials test the rule against real task outcomes. Validation promotes it when results improve, retires it when they do not, and continues guarding against regressions.
Trace evaluation
Each new trace is scored for task completion and coherence, while the trajectory is watched for stalls or regressions. Suspected failures trigger deeper tool and argument checks, with optional planning and efficiency diagnostics.
Candidate rules are tested against replay or live evidence before promotion, then re-checked as the rule set evolves.
Prove it on the failure that triggered it.
When replay is available, the candidate rule is tested against recent failures and protected wins, then promoted only when outcomes improve without regression.
Prove it live when replay is unavailable.
The candidate is trialed across matching live sessions and promoted only after its breach rate reaches zero or clears its baseline by a meaningful margin.
Keep the gate honest as rules accumulate.
Captured failures and protected wins are replayed against the current rule set, while labeled outcomes keep breach thresholds measured rather than guessed.
Let evidence decide what your agents learn.
Start buildingFirst-class adapters for six agent frameworks, a lightweight path for custom loops, and four optional read-only tools for consulting learned rules.
Your agents can list, search, and read learned rules or check a rule's validation status. Repair and administration stay outside their tool surface.
Keep your loop. Add the Repair Harness.
View integrationsInstall the package, connect your project, and wrap one turn boundary—no framework migration required.
Add the Python package to your agent environment.
Connect the Harness to the PandaProbe project receiving your traces.
Use a framework adapter or wrap one turn boundary in your existing loop.
Wire up your first repair loop.
Read the quickstartEverything you need to know about building self-repairing agents with the PandaProbe Harness.
Wire the Repair Harness into your agent loop and turn failures into validated rules the next turn can use.
Peak task-success uplift
+20%
Build your repair loop in the open.
Start building
