Back to workTrace forensics · Private build

AgentRewind

AgentRewind is my private build for replaying completed agent traces and comparing decisions, tool calls, and state changes. I describe its design here; no public repository or test output is available.

Trace import checks

I treat imported trace files as untrusted input and validate them before replay. These are design details for my private build; there are no public tests to inspect.

Signed imports with JWT verification; unsigned or signature-mismatched payloads are refused.
Payload size double-check: declared size is verified against actual size before parsing.
MAX_STEPS limits the replay length.
Input sanitization on every field before it hits the diff or replay path.
The importer detects five trace formats. Its parsing behavior is not publicly verifiable.

Diff engine

I compare recorded runs step by step to locate differences in decisions, tool calls, and state changes. The result depends on the data in the traces.

Divergence points: the first step where run A and run B disagree.
Latency deltas per step and across the full run.
Input and output token-count differences support cost regression checks.

Why privateAgentRewind is my private build. There is no public repository, demo, or test output, and I have not promised a release date.

Discuss your workflow with TechTide AI.

TechTide AI reviews AI workflows and Airtable systems, then builds the agreed fixes. I keep code and resources here; current service scope, prices, and terms are on techtideai.io. An Airtable review can recommend staying, repairing, connecting tools, or migrating.

Systems Decision Audit Airtable Rescue Decision Sprint alex@techtideai.io

Also on TechTide AIProduction Block Call Critical System Implementation

alexcinovoj.dev