Problem
Teams often cannot explain agent runs, drift, or token spend after an incident or surprising output.
After the run: timelines you can read, drift you can act on, and replay that does not oversell determinism.
Project information
- Category: Agent observability · Drift · Post-run evidence
- Focus: Timelines, release comparison, honest replay reporting
- For: Teams past “it worked in the demo”
- Role: Reference system behind reliability engagements
Overview
When an agent fails in production, teams need a record they can trust. Not a black box. Not marketing that pretends every “replay” re-runs the world identically.
Agent Lens gives operators readable run history, drift signals before the next promote, and analysis grounded in what was observed.
Work with me:
Mechanism
Trace timelines, drift checks, and replay analysis work from recorded evidence instead of silently re-running models.
What It Proves
Post-run review can diagnose behavior and cost without creating a second hidden model run.
Engagement Relevance
Useful for audits and retainers where operators need stable evidence for failure review and monitoring decisions.
CTA
Use this pattern to inspect whether current traces are enough to trust production agents.
Send a problem briefTechnical evidence
Frozen run trace and trace-grounded replay report (reference system example).
model_call planner Fix the auth token expiry bug in src/auth.py
tool_call read_file src/auth.py
tool_call read_file tests/test_auth.py
model_call coder Fix the expiry comparison to use UTC timestamps
tool_call write_file src/auth.py
replay_mode: trace_grounded_report
trace_digest: sha256:8f3a…c21eKnown limitation: Default replay analyzes recorded evidence; bit-identical live LLM re-execution is a separate, gated mode and is not the default production story.
Goal
Make agent behavior inspectable after the fact and cautious before the next release. Timelines operators can read. Drift awareness. Honest labeling of what replay means.
What it does
- Run timelines: a clear path through model, tool, and gate-style decisions for a given run.
- Drift awareness: surfaces that help teams notice when behavior shifts across releases.
- Replay reporting: analysis grounded in observed traces; live re-execution is a separate, gated mode when used at all.
- Operator evidence: built for incident review and reliability conversations, not vanity metrics alone.
Honesty by design
Default replay is framed as analysis of what already happened. That keeps buyers and operators from being misled about determinism.
- Trace-first storytelling: freeze and explain the observed path before claiming re-execution.
- Cautious promotion signals: drift views prompt judgment; they do not manufacture false confidence.
- Shared reliability narrative: cost blocks and human gates can appear in the same evidence trail as model steps.
Where it fits
- Use cases: production agent fleets, pre-deploy checks, incident postmortems.
- Signal: especially useful when demo success is no longer enough.
What I built
I built Agent Lens as an evidence product: visibility, drift caution, and honest replay framing. Observability that supports trust, not another opaque dashboard.
Why it matters for clients
Agent systems become operable when post-run evidence and release caution get the same seriousness as the agent itself. That is what audits surface and pilots install.