Ask before the agent acts
Agentic registers AI agents, checks each governed request against policy, routes exceptions to a human reviewer and records tamper-evident evidence
Try every flow safely with the built in procurement simulator
Procurement assistant asks to buy
$11,200.00
from conveyor-works
- Policy checked, above the $10,000 threshold
- Budget held, sent to finance reviewers
- Approved by an independent reviewer
- Simulated purchase recorded
Sample from the fictional demonstration data, simulated dollars
Agent registered · Key issued · Request allowed · Review requested · Request approved · Permit issued · Purchase recorded · Checkpoint signed ·
Three controls behind one call
The agent asks once. Agentic answers from the registry, the policy and the budget, then writes the record.
Registry
Every agent has an owner, a department, a lifecycle state and keys that expire. A suspended or retired agent is denied.
Authorization
A versioned policy answers allow, deny or needs review. An allowed request gets a single use permit bound to that exact action.
Evidence
Each decision, review and ledger change joins a hash chain with signed checkpoints, so a later edit can be detected.
Five flows in the pilot
Each one runs end to end in the simulator, from the request to the ledger.
A small purchase is allowed
The request is inside policy and budget. The agent receives a permit that works once, for that exact purchase, for 60 seconds.
Decision, permit and ledger entry recorded
A large purchase waits for a person
Above the approval threshold, the amount is held against the budget and routed to an independent reviewer, a different person from the one who started it.
Approved or rejected with a recorded reason
A blocked vendor is denied
Deny rules take precedence over allow rules, so a blocked vendor is stopped before any permit is issued.
Denial recorded with its reason
A request over budget is denied
Open requests reserve budget before anything runs, so every approval is backed by dollars that are actually available.
Budget counters reconcile with the ledger
An unknown outcome is reconciled
When the executor times out, Agentic keeps the budget held and reconciles the execution against the executor's record before committing or releasing it.
Committed or released after reconciliation
Connect an agent in four steps
The contract is plain HTTP. A TypeScript SDK ships in the repository and the simulator makes the same calls without code.
1
Register the agent
Give it an owner, a department and the actions it needs.
2
Issue a key
The key is shown once, carries scopes and expires.
3
Ask before acting
Call authorize ahead of each governed action and respect the answer.
4
Send activity
Heartbeats and run events make the agent visible on the overview.
Built today, and what comes next
Point at a tile, or tap it, to see the roadmap for each control.
Registry
Built today
Manual onboarding, owners, lifecycle states, tags and key rotation
Monitoring
Built today
Typed events, heartbeats, run outcomes and declared costs
Authorization
Built today
A versioned policy, deny by default and single use permits
Approvals
Built today
Independent reviewers, expiry, rejection and cancellation
Finance
Built today
US dollars in whole cents, reservations, a ledger and hard ceilings
Audit
Built today
An append-only log, a hash chain and signed checkpoints
The targets we build to
Every release is measured against these objectives.
300 ms
Authorization decision at the 95th percentile
99.9 percent
Monthly availability objective for valid decisions
60 s
From accepted telemetry to the dashboard, 95th percentile
100 percent
Of required transitions with an audit entry in release tests
Questions
1 of 7
How does Agentic handle money
Every purchase runs through a procurement executor that only accepts a valid single use permit. The pilot uses a simulated executor and records each purchase in a ledger, so you can run real policies end to end safely.
An owner registers each agent with a department, the actions it needs and a scoped key that expires. Registered agents report heartbeats and run events, so every one of them shows up live on the overview.
The agent calls authorize before each governed action. Agentic answers allow, deny or needs review, and an allowed request receives a single use permit bound to that exact action, which the executor checks before anything runs.
A member of the reviewer group named in the policy, and always a different person from the one who started the request. Reviewers see the exact action hash and approve or reject it with a recorded reason.
Every decision, review and ledger change joins a hash chain, and signed checkpoints anchor it. Any later edit is detectable, and auditors can verify the whole chain from the console.
With email and password, authenticator app codes for multi factor sign in, or single sign on through any OpenID Connect provider. Admins can require multi factor sign in for the whole organization, and the demo lets you try every role in one click.
Through a plain HTTP API with an OpenAPI description and a TypeScript SDK. Signed webhooks deliver events to your endpoints with automatic retries, and a background worker keeps approvals, permits and budgets current every few minutes.
See it decide
Sign in as an admin, an operator, a reviewer or an auditor in one click and watch every decision happen in real time.