Store event lineage by default
Plans, tool calls, observations, digests, and derived facts point back to their source events, so the memory graph stays auditable.
Agent run recorder
Find the break in every agent run.
TraceBreak captures memory, execution, and replay as one linked event lineage, so teams can debug AI behavior with evidence instead of screenshots and guesswork.
- Memory
- Every decision is stored with the evidence and source events that shaped it.
- Runs
- The runtime writes the audit trail as work happens, not after the incident.
- Replay
- Bad runs become inspectable, diffable, and repeatable paths.
Why it exists
Observability sees symptoms. TraceBreak shows the decision chain.
Agent failures usually hide in the space between a prompt, a tool result, and the next model decision. TraceBreak makes that space a first-class trace.
Keep the loop small and visible
The runtime records each step, each tool result, and each stop reason without turning the agent into a black-box framework.
Compare counterfactual behavior
Replay the same chain against a new model, prompt, or state and see the first step where behavior changes.
Live product shape
Pick a failure mode. See the break.
This is the core workflow: choose the run that failed, follow the event lineage, and compare the replay diff at the first changed decision.
tracebreak replayrun_8f31
01 start
Goal enters runtimeSummarise a customer refund request and choose the next action.
02 memory
Evidence attachedThree source events link the policy, account state, and customer history.
03 break
Prompt drift detectedThe replay used the same event lineage and isolated the first plan that changed after the policy prompt update.
- action: approve_refund
+ action: request_manager_reviewOne product
No separate projects for the end user.
TraceBreak folds memory, runtime, and replay into one operational surface. The implementation can evolve underneath; the team debugging the run sees one product and one vocabulary.
Built for operators
Use it when the run matters.
TraceBreak is for teams shipping agentic workflows where a bad decision has to be explained, reproduced, and fixed.
Capture
Record each run as typed events with source links and tenant isolation.
Inspect
Open a failed run and walk the decision tree from output back to evidence.
Replay
Run the same chain against a new model, prompt, or state snapshot.
Ship
Compare diffs, confirm the fix, and keep the trace attached to the incident.
Private beta
Bring one painful run.
The fastest demo is a real failure: a hallucinated tool, a prompt drift, a loop, or a model swap that changed behavior.