Black Box
The flight recorder for AI agents

Prove exactly what your AI agent did.

Black Box records every model and tool call your agent makes, from outside the agent, in a tamper-evident log anyone can verify. Built for teams shipping agents, and the insurers covering them.

Book a 15-min callTry the tamper test

Free 2-week pilot on one workflow. No card until it ends.

run-0193.ledgersample
  1. 1
  2. 2
  3. 3
verifying…
I build agents

Debug runs, and hand customers proof of what the agent did.

See how it works
I insure agents

Verify claims and spot fraud with an independent record of what the agent did.

For insurers

When an agent goes wrong, can you prove what happened?

01

Scattered logs

Model calls in one tool, shell commands in another, API calls somewhere else. Reconstructing one run means stitching them together by hand.

02

The agent's own trace

If the agent reports what it did, you're trusting the thing under review. A wrong or incomplete self-report looks identical to a right one.

03

Logs anyone can edit

Plain log files and database rows can be changed by anyone with access, and nothing shows that they were.

Live tamper test

Edit the record. Watch the chain break.

A sample agent run, hashed with SHA-256 in your browser. Click any row, change an argument, and the verifier points at the exact entry.

ledger · sample run
  1. 01
    plan
    …
  2. 02
    vault.read
    …
  3. 03
    sql.query
    …
  4. 04
    decide
    …
  5. 05
    vault.write
    …
  6. 06
    shell.exec
    …
  7. 07
    http.get
    …
  8. 08
    summarise
    …
  9. 09
    slack.post
    …
  10. 10
    ticket.close
    …
verifier · illustrative output
$ bbx verify run-0193.ledger
…

Real SHA-256, computed in your browser. Nothing is sent anywhere. The sample run is illustrative, not customer data. The offline verifier is in pilot.

The agent never writes its own record, so it can't edit it.

  1. 01
    Your agent

    Unchanged. Its model and tool calls are pointed at the gateway.

  2. 02
    Gateway

    Sits outside the agent and records every model and tool call as it passes through.

  3. 03
    Ledger

    Each entry is hashed together with the one before it. Edit one and every later hash breaks.

  4. 04
    Verifier

    Checks an exported ledger offline, built so a third party can run it.

  5. 05
    Replay

    Step through exactly what was called, when, with what result.

Architecture and the ledger format →

Step through exactly what happened.

replay · run-019314:02:11.000Z
model call
plan
args
task=rotate staging db credentials
result
3 steps planned

Tracing tells you what happened. Black Box gives you evidence you can hand to someone else.

Tracing tools and LLM gateways are good at what they do, and you'll likely keep them. Black Box answers a different question.

Question answered

Tracing tools: What did my agent do?

Black Box: Can I prove what my agent did?

Where it records

Tracing tools: Inside your code, via SDK or callbacks

Black Box: At the gateway, outside the agent

Edits to the record

Tracing tools: Possible, and not detectable

Black Box: Break the hash chain

Checking it later

Tracing tools: Trust the vendor's dashboard

Black Box: Verify an exported ledger offline

Insurers: settle claims from a record nobody can quietly edit.

Attach Black Box to a policyholder's agent and get an independent, tamper-evident record of what it did, so you can verify claims and spot fraud.

How it works for insurers
Pilot terms

How the pilot works

Free 2-week pilot

If you keep using it after 2 weeks, it becomes a monthly subscription at a price we agree before the pilot starts. We don't take a card until then.

Book a 15-min call
  1. 115-min call: pick one agent workflow.
  2. 2We route it through the Black Box gateway with you (a few lines of config).
  3. 3For two weeks you get a full, verifiable record of every model and tool call.
  4. 4In return: honest feedback, and a testimonial if it earned one.
  5. 5Keep it after two weeks on a monthly subscription. No card until then.

Questions teams ask us

Do you have customers?

Not yet. We're running free 2-week pilots with a small number of teams, and you'd be one of the first.

What happens after the 2 weeks?

If you keep using it, it becomes a monthly subscription at a price we agree before the pilot starts. If not, you walk away.

Do I need to give you my card to start?

No. We don't take a card until the pilot ends.

How big is the setup?

One workflow and a few lines of config to route its calls through the gateway. We do it with you on the pilot.

Will it slow my agent down?

The gateway adds a hop. We'll measure it on your workflow during the pilot and show you the numbers rather than quote one here.

Which stacks do you support?

We confirm that on the call, against the workflow you want to record.

Book a 15-min call. Bring one agent workflow and we'll show you what its record would look like.

  • Setup
    One workflow, a few lines of config
  • Cost to start
    Free for two weeks, no card
  • Stacks
    Confirmed on the call
  • Your data
    Region and retention agreed in writing
Pick a time

15 minutes, no prep needed.

Not ready to talk? Get the 2-page technical brief.