Skip to content

OPERATE WITH CONFIDENCE

Run agents in production.

You can put agents on real work because you can see, and allow, everything they do. Reads flow; writes wait for a person. Scope is enforced at the database. And every action lands in a tamper-evident record you can export.

THE CHECKPOINT

Reads flow. Writes wait for you.

An agent reads freely, but the moment it wants to change something the turn suspends and asks, in plain language, not a diff. Approve it once, approve it for the whole session, or deny it. A denial is not a dead end: it steers the agent toward another approach.

CHECKPOINT · APPROVAL NEEDED
turn 6 · suspended
Allow user__create to run? 1:48

The assistant is asking permission to create user “m.torres@acme.co” with the role “analyst”.

user__create · role=analyst · tenant=acme
Approve Approve for session Deny

Denying steers the agent; it adapts and proposes another path.

Plan mode
PLAN MODE · read-only

Tool catalog

records__list
records__get
schema__describe
audit__read
Write tools Not in catalog until the plan is approved

PLAN MODE

The write tools are not blocked. They are absent.

In plan mode the agent literally does not have write tools in its catalog; it can only read and propose. Nothing to bypass, because there is nothing there. The write surface appears only after a person presses “Begin implementation”.

SCOPE BY CONSTRUCTION

An agent can only touch what it is scoped to.

Scope is not a prompt instruction the model can talk its way around; it is enforced below the agent, in the platform.

Tenant isolation at the database

Row-level security enforces the tenant boundary in Postgres itself. No query, human or agent, crosses it.

Per-view tool surfaces

A view-bound agent gets exactly the actions that view exposes: the buttons a human would click, and nothing more.

One-record pinning

A per-record agent is pinned to the single record you have open. It cannot reach sideways to another.

The safety floor refuses the rest

A cross-tenant or cross-record argument is refused at the policy floor before it runs, and the refusal is recorded as a blocked row in the log, so an attempt is as visible as a success.

DECISION LOG
tenant · acme
TimeActionPolicyOutcome
14:02:11 records__list reads-approve allowed
14:02:36 user__create writes-ask approved
14:03:07 records__get·x-tenant safety-floor refused
TAMPER-EVIDENT · SIGNED RECEIPTS

THE RECORD

Every action, hash-chained and signed.

Each tool invocation becomes one row in a per-tenant SHA-256 hash chain, carrying a signed decision receipt: who, what, which policy, and the outcome. The log is filterable and exports to JSON or CSV, a tamper-evident record of everything an agent did, and everything it was refused.

  • A signed decision receipt on every row: principal, tool, policy, and outcome.
  • Hash-chained per tenant, so the record is tamper-evident.
  • Filter the log and export it to JSON or CSV.

Roadmap On the roadmap: a portable Evidence File export you can hand to a third party.

APPROVE ANYWHERE

One Checkpoint. Three places to answer it.

The same approval reaches you wherever you are: the web rail beside the agent, the ticket timeline, or your phone. The decision is the same; the place is yours to choose.

9:41
Ops agent Online
I drafted the end-of-shift summary email to the ops team.
Ready to send. This one needs your approval.
send_email executed

Approved

send_email · confirmed on device

Web

The web rail

The approval card sits right next to the agent transcript, with the plain-language explanation inline.

Tickets

The ticket timeline

When an agent works a ticket, its plan-mode diagnosis and approvals live on the ticket itself.

Mobile

Your phone

The chat input transforms into the approval card with masked arguments and a countdown; approve on the go.

On mobile, sensitive tools can require Face ID or a fingerprint before they run, an on-device biometric gate on the approval itself.

EVAL OBSERVER · LIVE
run · centinela-3

Verdict

PASS

0.94

score

20 / 20

Gate groups

  • Scope respected
  • Approvals honored
  • No cross-tenant reach
  • Grounded answers

GRADED BEFORE THEY SHIP

You grade the agent, and see the rubric.

An eval harness runs your agents against 20 gate groups, and an Agent-as-a-Judge returns a Verdict with a per-criterion rubric. A live observer streams the judge’s reasoning as it scores, so a passing agent is a shown result, not a claim.

DEPLOYMENT POSTURE

Where it runs, and who it answers to.

Your cluster or ours
Deploy into your own Kubernetes, or run it hosted. Same platform either way.
Keycloak SSO
Authenticate against Keycloak, and the roles you already manage map straight through.
Multi-tenant by construction
Tenant isolation is enforced at the database, not bolted on per feature.

OPERATE WITH CONFIDENCE

See everything. Allow what should happen.

Put agents on production work with a Checkpoint on every write, scope enforced at the database, and a tamper-evident record you can export.