rokan.

The AI forward-deployed engineer for teams with agents live at customers

Keep your agent working at every customer.

Your agent tells a caller "I'm transferring you now." No transfer happens. Rokan catches that, says why, writes the fix and proves it on every customer's past conversations before it ships. So the next caller actually gets through.

160 of 168 broken conversations caught · 1 false alarm in 65 · on a telecom support benchmark see the test and its limits

deploy.rokan.computer/demo Open

Loading the live product. Open it in a new tab

The product itself, not a recording. Made-up data; nothing is saved.

Reads the runs your agent already produces

From one broken run to a fix you can trust.

Rokan treats every customer as its own setup. A fix only ships if it helps the customer that broke and breaks no one else.

  1. 01

    Reads the runs

    Every new conversation, in the format your stack already exports.

  2. 02

    Finds why

    In that customer's setup, in plain words, with the lines that show it.

  3. 03

    Tests every fix

    Replayed on every customer's past runs. A fix that breaks one is blocked.

  4. 04

    Waits for you

    A small prompt or config change. Nothing reaches your agent until you approve it.

It also catches runs that looked fine.

The failures your customers notice before you do.

Tested with builder and grader kept apart.

Two results on a public benchmark (tau2-bench, telecom support). Our code was frozen first; our co-founder's side drew the fresh tasks and scored every result.

160 of 168

Failed conversations caught

On 233 fresh conversations with simulated customers, with 1 false alarm in 65 good ones. Frontier models reading the same transcripts unaided caught fewer; given the same checklist, GPT-6.2 matched it. The edge is knowing what to check.

19 → 41 of 50

Tasks solved after the fix

On one benchmark, in all three attempts, Rokan's verified fix about doubled passes over the unchanged agent (19 to 41, 21 to 43, 21 to 44), and beat the same model proposing fixes without Rokan.

FIG. 01 · BROKEN CONVERSATIONS CAUGHT, OF 168At most 5% false alarms on the 65 good ones
Rokan's checks
1601 false alarm in 65
Fable 5.2, reading the same transcripts
1066 false alarms in 65 · over the 5% limit
GPT-6.2 Astra, same transcripts
151no setting met the 5% false-alarm limit; best shown: 7 to 9 false alarms in 65
tau2-bench, 233 fresh conversations with simulated customers (168 failed, 65 fine). One domain; the checks were built by hand for it. Thresholds were set on a separate half first and applied once; our co-founder's side drew the tasks and scored every result. Given the same checklist in its prompt, GPT-6.2 matched the checks exactly. Rokan vs Fable 5.2: paired p < 0.001. A benchmark, not a customer result.
FIG. 02 · TASKS SOLVED, OF 50Unchanged agent With Rokan's verified fix
Attempt 1
19 41
Attempt 2
21 43
Attempt 3
21 44
02550
tau2-bench telecom, 50 held-out tasks, simulated customers. Our code was frozen before the run; our co-founder's side drew the tasks, ran and scored every attempt. One domain: a benchmark result, not a customer result.

One domain, simulated customers. Benchmark results, not customer results.

You control where your data goes.

Two ways to run it, and one rule for both. Pick the way your customers' contracts allow.

In the beta

You upload runs. Names, phone numbers, emails and card numbers are removed the moment they arrive, and everything is deleted within 30 days, or the same day you ask.

Strict mode

Runs on your machine or cloud with a model you choose. Rokan refuses every outside connection, and you can run it fully offline. Available under contract. Read the one-page security note

Nothing ships on its own

Fixes arrive as a prompt or config change. A person on your team approves each one, and can undo it. We never train on your data.

Free for solo builders. 14 days free for teams.

Upload 20 to 200 runs and in minutes you get the report and a tested fix. Nothing to install, no keys, no card. We take teams in a few at a time.

We're early: you'd be one of the first teams.