The Agent Decision Record
Proof of what your agents decided
An Agent Decision Record is a signed, append-only record of every decision an AI agent makes: who acted, under which rule and inputs, who approved, and what happened, verifiable offline by anyone. Tenbri is the open-source one. One line for Claude, OpenAI and LangGraph agents; plain files on storage you control.
$ pip install --pre tenbri
import tenbri
tenbri.init() # TENBRI_STREAM, TENBRI_AGENT, TENBRI_APPROVER
tenbri.instrument_claude_agent_sdk() # every tool call becomes a signed decision
js
import { init } from 'tenbri'
import { instrumentLangChain } from '@tenbri/langgraph'
init() // TENBRI_STREAM, TENBRI_AGENT, TENBRI_APPROVER
instrumentLangChain() // every tool call becomes a signed decision
// or: instrumentClaudeAgentSdk(options) from '@tenbri/claude-agent-sdk'
// or: instrumentOpenAIAgents() from '@tenbri/openai-agents'
python
import tenbri
tenbri.init()
@tenbri.decision("issue_refund", target="stripe:{charge}", policy="policy:refund@14")
def refund(charge: str, amount: int) -> None:
if amount > 500:
raise tenbri.Refused("over the ceiling") # recorded as refused
stripe.Refund.create(charge=charge, amount=amount)
refund("ch_81f2", 4200) # intent before it runs, outcome after
js
import { Refused, decision, init } from 'tenbri'
init()
const refund = decision(
{ action: 'issue_refund', target: 'stripe:{charge}', policy: 'policy:refund@14' },
async ({ charge, amount }: { charge: string; amount: number }) => {
if (amount > 500) throw new Refused('over the ceiling') // recorded as refused
await stripe.refunds.create({ charge, amount })
},
)
await refund({ charge: 'ch_81f2', amount: 4200 }) // intent before, outcome after
Tenbri UI
See what your agents decided #
tenbrid serves a read-only UI over your streams: dashboards, every decision with its outcome, traces as timelines, and a verification check. Open it with a bearer token; it cannot write.
Overview, then the errors, one decision's timeline and signed chain, its whole trace, and the integrity check.
See it run
Record, verify, tamper, get caught #
A refund function decorated with @tenbri.decision, called three times. Then someone edits one byte on disk. Real output from a run, with hashes and paths shortened.
- 1Record
- 2Read
- 3Verify
- 4Tamper
- 5Caught
From the blog · Incident analysis
Five days to “it was us” #
In July 2026, OpenAI agents broke out of an evaluation sandbox and attacked Hugging Face. Attribution took five days. What a signed record of every agent decision would have changed, and what it would not.
- ~17,600 agent actions Hugging Face recovered from its own logs
- 5 days from the public disclosure to “it was us”
- 1 query with a Tenbri record:
tenbri query --target-domain, on the roadmap
What you get
Built for the day something goes wrong #
Disputes, incidents, policy changes and audits all ask the same question: what did the agent decide, and on whose authority?
$ tenbri verify
verified 8 entries, stream tnb1:986d8e32…3588, pinned to sha256:986d8e32… (the pin recorded in ~/demo/keys/pins)
$ tenbri verify # after someone edits one batch on disk
broken at seq 5: bad_signature: the signature does not verify
4 entries verified before it, stream tnb1:986d8e32…3588
$ tenbri verify audit/support --key sha256:986d8e32… # an auditor on another machine pins it
Anyone can check it, offline
Hand an auditor the stream and one fingerprint. Verification checks canonical bytes, the hash chain, every Ed25519 signature and every key, with no access to your systems. A checkpoint they kept catches truncation and forks.
$ tenbri query --rule policy:refund@14
3 2026-10-20T16:49:42.442Z decision intent agent:support issue_refund stripe:ch_3PqX - verified
5 2026-10-20T16:49:42.485Z decision intent agent:support issue_refund stripe:ch_9ZzY - verified
7 2026-10-20T16:49:42.525Z decision intent agent:support issue_refund stripe:ch_0 - verified
$ tenbri query --rule policy:refund@14 --json | jq .entry.action # JSON when piped
Every decision made under the old rule
Inputs are pinned by reference, version and keyed digest. When a policy changes, list what was decided under the previous version and review it. Each answer is re-checked against storage before it is returned.
unknown_actions = "reject"
[actions.issue_refund]
target = "stripe:{charge}" # filled from the tool call's arguments
policy = "policy:refund@14" # the rule that allowed it
approval = true # a person must approve first
inputs = ["policy:refund"]
[actions.issue_refund.fields."ext.refund.amount"]
required = true
type = "integer"
Say what each action must record
Declare actions once in tenbri.toml. The SDKs read it to map tool calls to actions, targets and policies, tenbrid enforces it at write time, and CI checks whole streams against it.
$ tenbri receipt create ./audit/support --entry 3 --out refund-3.receipt
$ tenbri receipt verify refund-3.receipt --key sha256:9c1e…
proves seq 3 (decision intent), id 01M4XBH0Y6…, ts 2026-10-14T14:02:13Z (claimed by the writer)
in stream tnb1:1702…b7d8, checkpoint at size 8, pinned to sha256:9c1e…; key history: 2 key entries
Proof you can hand over
A receipt proves specific entries with their inclusion proofs, the key history and an optional RFC 3161 timestamp. A customer, a regulator or the police verify it alone, without seeing the rest of the stream.
$ tenbri erase ./audit/support --subject subj_7f3c91d2 --reason dsr:2026-1042 --yes
erased subject subj_7f3c91d2: 2 data key epochs deleted; erasure recorded at seq 7
$ tenbri verify ./audit/support --key sha256:9c1e…
verified 8 entries, stream tnb1:1702…b7d8, pinned to sha256:9c1e…
Erase a person without breaking the chain
Only hashes go in the chain. Bodies are encrypted and input digests are keyed per subject, so deleting a subject's keys makes their data unreadable while every chain, checkpoint and receipt still verifies.
$ export TENBRI_STREAM=r2://audit/agents/support CLOUDFLARE_ACCOUNT_ID=…
$ curl -X POST https://tenbrid.internal:7070/v1/streams/support/intents \
-H "Authorization: Bearer $TOKEN" -H "Idempotency-Key: $(ulid)" \
-d '{"action":{"type":"issue_refund","target":"stripe:ch_81f2"}}'
{"entry":{"kind":"decision","phase":"intent","seq":3,…},"replayed":false}
Your storage, any language
A stream is plain files in a directory, S3, R2 or MinIO, with the keys kept apart. Write from Python, TypeScript or Rust, or run tenbrid and use REST or gRPC from anything else.
A new kind of record
Logs, traces, decision records #
Observability tools such as Langfuse and MLflow answer how an agent ran. A decision record proves what it decided and on whose authority. Keep both: Tenbri links each decision to its OpenTelemetry trace.
| Question | Logs | Traces | Decision records |
|---|---|---|---|
| What it answers | What a service printed | How the agent ran: steps, latency, tokens, cost | What the agent decided, and on whose authority |
| Who it serves | Operators | Engineers improving the agent | Auditors, customers, regulators and incident review |
| Can it be altered? | Yes: edited, rotated, deleted | Yes: sampled, rewritten, expired | No: any change fails verification |
| Can a third party verify it? | No, they must trust you | No, they need your account | Yes, offline, with one fingerprint |
Architecture
Write, sign and chain, store, prove #
One Rust core behind every SDK, the CLI and tenbrid. Turn on the optional parts when you need them; the rest is on the roadmap.
Records these frameworks ask for
Evidence, not screenshots #
Tenbri gives you the records. It does not make a system compliant on its own.
- EU AI Act Art. 12, 14, 19, 26
- GDPR Art. 22
- DORA
- NIS2
- SOC 2
- ISO/IEC 27001
- ISO/IEC 42001
- NIST AI RMF
Start recording decisions in five minutes #
Open source, Apache-2.0, written in Rust. Plain files on storage you control, and no telemetry.