Agent observability

Your agents, from prompt to merge.

Weave traces every session your coding agents run. See where they loop, what they spend, and what ships.

Book a demoRead the research

Sample sessions

  1. 01promptFix rounding in invoice totals
  2. 02model.call18.4k tokens · $0.21
  3. 03tool.readbilling/invoice.ts
  4. 04tool.editinvoice.ts · 3rd edit
  5. 05tool.testbilling.spec.ts · 2 failedFailed
  6. 06tool.editinvoice.ts · 6th editFlagged
  7. 07promptUse banker’s rounding instead
  8. 08tool.testbilling.spec.ts · passed
  9. 09pull_request#1842 merged

9 spans · Illustrative, not a customer session.

See the full trace

What you see

Every session, in one place.

Session timeline

Every span on one clock, from prompt to pull request.

Loop patterns

Thrash, error streaks, and exact repeats, flagged by pattern.

Spend to outcome

Session cost, joined to the work that merged.

Agent by agent

Interactive and background agents, side by side.

Readable by agents

Weave’s MCP server lets connected agents query your team’s metrics.

Cost coverage

Per-user cost attribution for

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Codex
  • Devin
  • Windsurf
  • Amazon Q
  • Sourcegraph
  • Augment
  • Kiro
  • Firebender

Illustrative shapes

Session trace

Every step, on one clock.

Prompts, model calls, tool calls, tests, and the pull request. Each with its time, tokens, and cost.

Session 4f2a · fix invoice rounding

Claude Code · 12 spans · 38 min · 69.2k tokens · $0.70

  • Prompt
  • Model call
  • Tool call
  • Change
Sample session, illustrative

Autonomy isn’t success

Running isn’t landing.

Most sessions run without a correction. Very few land on the first prompt.

93%ran without a correction
3.5%landed on the first prompt

3.7 prompts per session, on average

Weave AI Impact Report, Q2 2026. 79,518 agent sessions; rates use classified outcomes, and about 41% are unclassified.

Read the report

What Weave tracks

Six signals. Six denominators.

One number can’t describe an agent.

  1. 01

    Autonomy

    Runs without a correction or an interrupted tool call.

  2. 02

    First-pass success

    Accepted without another prompt or a rewrite.

  3. 03

    Accepted output

    Agent changes that are reviewed and merged.

  4. 04

    Delivery

    Task start to accepted change.

  5. 05

    Quality

    Reverts and rework after merge.

  6. 06

    Cost

    Spend per accepted change.

The syntax era of agent reliability is over. The behavioral era is just getting instrumented.
From “Detecting doom loops,” Weave blog

Loops and stalls

Catch the stuck session early.

Bad tool calls are rare now. What’s left is the right call, made correctly, for a plan that stopped working.

  • Same-file thrash

    Five or more edits to one file, no progress.

  • Error streak

    Three or more failed tool calls in a row.

  • Exact repeat

    The same call, same arguments, again.

spirals in about 25,500 sessions
582
worst invalid-argument rate of any model
0.04%

Weave Router, 30 days: about 275,000 requests and 150,000+ tool calls. Detectors ran in shadow mode.

What 582 spirals taught us

Spend to outcome

Spend, joined to what merged.

Weave joins traces to cost and source control.

$26of every $100 in session spend was tied to work that merged

Weave AI Impact Report, Q2 2026. 21,907 sessions with trace and cost data. A share of spend, not a savings figure.

See Token intelligence

Bring your agent sessions into view.

Right beside the output, quality, and cost Weave already measures.

Book a demoStart for free