Build agents that keep self-improving.

Ratel finds problems, fixes them and runs evals on your agent, for higher accuracy, lower latency and no wasted tokens, turn after turn.

Get started ↗See how it works

Backed by

  • Italian Founders Fund
  • Angel Invest
  • Vento
  • IAG

Trusted by developers at

Poelis, Kestral, Pillar, Cato, Qura, GP+, Deeptree, Vibiz

How it works

From real runs to better agents.

Ratel reads the traces you already have, then finds, fixes, tests and ships changes. Every new run feeds the next one.

Your traces

  • OpenTelemetry
  • Langfuse
  • LangSmith
  • PostHog
  • Postgres
  • LangGraph
01 Find See relevant problems first. 02 Fix Get recommended changes. 03 Evals Test on your production data. 04 Ship Ship via SDK or your coding agent.

Human approval

or automatic

Your agent, in production+ Ratel SDK · optional

New runs feed the next round

Find what you can improve.

See all the tasks inferred from your users, and spot which ones need a fix and why.

Agent health · last 7 days

First try

75%

of 632 turns

Junk in lists

76%

hits rank low

Wasted calls

45.5

per 100 turns

What needs a fix

4 of 18 issues · % of turns

  • Lists are padded

    Waste · 76% of turns: The same tool, called repeatedly.

    76%
    Fix now

  • Wrong tool keeps showing up

    Wrong pick · search_hotels · 35% of turns: One tool tried, then another.

    35%
    Fix now

  • Agent keeps correcting itself

    Failures · 24% of turns: Requests fail partway.

    24%
    Watch

  • Refunds over $500 have no tool

    Missing pieces · 12% of turns: Requests no tool or workflow handles.

    12%
    Watch

Tool craft · 1 of 4

Tighten the tool catalogue.

4 changes
  • search_hotelssays when to use it, and when not
  • choose_seatreplaces get_seat_map + select_seat
  • hold_roomnew: hold a room while the traveller decides
  • legacy_export old_searchno calls in 30 days

Expected effect, in replay

23 → 28 of 32 searches right

Fix what matters, right away.

Ratel proactively recommends fixes that improve your agent’s cost, latency and accuracy.

Run evals before it goes live.

Replay past sessions with the fix on, and try different settings to find the best impact on accuracy, cost and latency.

Replay with

195 past sessions · 32 searches

Rewrite + Merge: accuracy 28 of 32, cost $8.43 per 1k sessions, p50 latency 3.2 seconds.

Accuracy

28of 32 searches

+16 pts · was 23

  • passed before
  • fixed by replay
  • still failing

Cost · per 1k sessions

−25%

−$2.77 · was $11.20

Before$11.20
After$8.43

≈ $2,770 saved per 1M sessions

Latency · median

3.2s

−0.9s · was 4.1s

Before4.1s
After3.2s

Ship changes straight to production.

Hand it to your coding agent, or let the Ratel SDK apply it live. New traces feed the next round.

Approved fix · tool craft

Tighten the tool catalogue

search_hotels+ 3 changes

In replay23 → 28 of 32

You approve, then pick a way in

  • ManualPaste the prompt into your coding agent.

    tools/catalog.ts +31 −58

  • MCPYour coding agent pulls it and opens a PR.

    PR #142

  • SDKApplied live. No redeploy.

    search_hotels v2LIVE

Production

Your agent,
one fix better.

Accuracy, last 7 days

New traces feed the next round

Deploy the SDK to save context and time.

Use runtime data to pick the right tools, keep context lean and adopt approved changes without a redeploy.

Explore the SDK

Adaptive ranking

Learns which tools actually get called.

why is the build broken?
  1. 1gh_run_list↑ learned0.94
  2. 2docker_build0.71
  3. 3slack_post_message0.18

Progressive disclosure

Each step sees only the tools it needs, not your whole catalogue.

Without Ratel · every step84 tools · 20.4k tokens
With Ratel · refund step5 tools · 1.1k tokens
  • find_booking
  • get_policy
  • refund_payment
  • notify_user
  • escalate_to_agent

Direct changes

Approved fixes go live. No redeploy.

find_account · versions

  1. v1Finds an account.revert
  2. v2+ by email or phonerevert
  3. v3+ call it before any billing toolLIVE

Every run feeds the next one.

Get startedRead the docs

FAQ

Before you start

01Do I need the SDK?

No. Start with OpenTelemetry traces.

02Can I bring traces from my observability tool?

Yes. Import them from the tools you already use, such as OpenTelemetry, Langfuse, LangSmith and PostHog.

03Who applies the changes?

Your team reviews and applies proposed changes.

04What data do I need?

LLM and tool-call spans. See the setup docs for supported fields and configuration.

Start with a first look together.

Book a setup call