Thought Needle — trajectory capture

Capture what your agents really do.

A trajectory is the full record of what an AI agent actually did — every observation, decision, and tool call on the way to a result. We capture those from the real world and turn them into environments, analytics, and signal.

Get in touch See how Harnest works

↳ PII is stripped in the browser, before any trajectory leaves the machine.

capturing haystack · 178 pts

1 trajectory found 0 PII kept

The material

Models are trained on the open internet. But the most valuable record — how real work actually gets done with AI — is scattered across terminals, IDEs, and dashboards, and it vanishes the moment a session ends.

We think that record is the raw material for the next generation of AI. Thought Needle exists to capture it well: faithfully enough to learn from, privately enough to trust.

Capture → clean → use

One pipeline,
three moves.

01 — Capture

The real thing, not a proxy

Trajectories from actual work — not synthetic benchmarks or scraped transcripts. What agents do when the task is real.

02 — Clean

Private by construction

PII is detected and stripped on the client, before capture. What we store is structure and behavior, not secrets.

03 — Use

Turned into leverage

RL environments, analytics, and hiring signal — built on trajectories that reflect how your teams truly operate.

The capture layer

Harnest records trajectories
without recording you.

Harnest sits alongside your agents and captures each run as a structured trajectory. Redaction happens client-side — names, keys, paths, and secrets are masked on the machine where the work happens. Only the cleaned trajectory is ever transmitted.

Client-side PII detection & masking, pre-transmission
Structured steps: observations, actions, tool calls, results
Deterministic redaction you can inspect and audit
You keep ownership of every trajectory you capture

How the privacy model works

step 04 · tool_callredacted · 2

{
  "action": "write_file",
  "path": "/Users/██████/api/keys.ts",
  "author": "████████",
  "token": "sk-████████████",
  "diff": "add retry to fetch()",
  "ms": 812
}

✓ masked on-device — never sent raw

Two ways teams use trajectories

One capture layer.
Two clear jobs.

We're early, and deliberately focused. Trajectories power both of these — pick the one that's yours.

For enterprises

Make your agents measurably better

See inside your agents' real trajectories, build custom RL environments from your own workflows, and get analytics on where AI is helping — and where it isn't yet.

Explore enterprise For recruiters

Find who's genuinely AI-native

We analyze coding-agent trajectories to see how candidates actually work with AI — how they steer, verify, and recover — so you can hire for real fluency, not résumé keywords.

Explore recruiting

Have trajectories
worth capturing?

We're working closely with a small number of early teams. If any of this is your problem, we'd like to hear about it.

[email protected]