Skip to content

Govern AI agentsControl every action. Prove every decision

Fairway evaluates each action an agent proposes, applies your policies and risk controls, and keeps evidence of the decision.

Overview

A snapshot of your agent fleet.

Last 24h
Events1,284Total executions
Active agents2Registered
Classifications312Flagged events
Alerts0Unresolved
Event activityHourly volume, last 24h
24 buckets
Recent eventsView all
AgentDecisionFrameworkWhen
payment-serviceApproveREST10/8/2026, 11:41:34 AM
payment-serviceApproveREST10/8/2026, 11:41:34 AM
payment-serviceBlockREST10/8/2026, 11:41:34 AM
payment-serviceApproveREST10/8/2026, 11:41:34 AM
cs-agent-send-supportEscalateLangGraph10/8/2026, 11:39:04 AM
cs-agent-send-supportApproveLangGraph10/8/2026, 11:39:03 AM
QuickstartWrap any agent. Every call is captured and classified.
Get an API key
1. Install the SDK
pip install agentrust-py
2. Run your agent
from agentrust_sdk import harness, embed_gateway

embed_gateway()  # starts local governance gateway

@harness(agent_id="my-agent")
def my_agent(user: str, input: str) -> dict:

Agents no longer just answer questions. They act

One routine support ticket, handled by an agent with direct access to your systems. Scroll to watch what it does, and why.

01 A ticket arrives

Friday, 6:42 pm. A customer writes in, and your agent answers alone.

Alex Morgan

Support agentSorry about that! Looking into it now.

No person reviews this ticket. The agent handles it end to end.

An illustration. Names and amounts are examples.

Every proposed action is checked before it runs

Fairway sits between the agent and the systems it acts on. It evaluates the action, decides the outcome and records why.

How Trust Runtime decides
DecisionsExample data
  • Refund a customer £4,800support-agentEscalateOver the £1,000 refund limit
  • Open a patient's medication historybilling-agentBlockRequester is not on the care team
  • Approve a £6,200 motor claimclaims-agentApproveWithin the £7,500 limit, documents present
  • Grant admin access to productionops-agentBlockOutside the agent's scope
  • Raise an oven setpoint to 205°Cline-agentEscalateOver the 190°C ceiling

Check against policy

Rules live in YAML packs a reviewer can read.

How Trust Runtime decides
Held for reviewEscalate

Refund a customer £4,800 for a cancelled order

  • The order exists and matches the customer
  • The refund is within the amount paid
  • Over this agent’s £1,000 refund limit
ReasonOver the agent’s limit. The order is valid.

Send risky actions to a person

The request, the failed checks and why it was held.

How review works
Decision recordExample data
Decision
Escalated, then approved
Rule
refund_limit
Reviewer
finance-lead
Reason
Over limit; order verified
Content hash
9f2c…a41e
Ledger hash
51d0…7bc3
Previous hash
c88e…02f9

Keep the reason for every decision

The rule, the inputs and who acted, with a hash.

How Trust Audit records

Different industries. The same check before every action

One agent action per industry, and what Fairway decides.

Discuss your use case

Financial services

Refunds stop at the agent's limit

A £4,800 refund is over the agent's £1,000 authority, so it waits for an approver.

Refund a customer £4,800 for a cancelled order. Checks: Order matches customer: yes; Within amount paid: yes; Within £1,000 authority: no; Permitted refund method: yes. Escalated: £4,800 is over the agent's £1,000 limit. Held for an approver; nothing is paid.

Healthcare

Records stay with the care team

Billing is not on the patient's care team, so the record stays locked.

Retrieve a patient's medication history to prepare a discharge summary, on behalf of a billing team. Checks: On the care team: no; Purpose fits the role: no; Only the data needed: yes; Recorded to the patient: yes. Blocked: Not on the care team. The record is not released; the privacy officer is told.

Insurance

Routine claims go straight through

All four checks pass and £6,200 is inside the £7,500 limit, so the payout may proceed.

Approve a £6,200 motor claim for payout. Checks: Policy active on loss date: yes; Documents complete: yes; Within £7,500 authority: yes; No fraud flags: yes. Approved to proceed: All four checks pass. The payout may proceed, and the basis is kept.

Enterprise operations

A ticket can't grant production access

Admin on a production database is outside the agent's scope, so nothing changes.

Grant administrator access to a production database for a contractor account. Checks: Requester entitled: no; Inside the agent's scope: no; Time- and scope-limited: no; Named approver recorded: no. Blocked: Outside the agent's scope. Nothing changes; the platform owner is told.

Manufacturing

Setpoints stay inside the spec

205°C is over the 190°C ceiling, so the oven holds at 180°C until an engineer decides.

Raise an oven temperature setpoint from 180°C to 205°C to speed up a batch. Checks: Inside 170–190°C: no; Reason supported: yes; Line can be changed: yes; Agent may change it: yes. Escalated: 205°C is over the 190°C ceiling. The setpoint stays at 180°C until an engineer decides.

Scenarios illustrate proposed product applications. Actions, amounts, thresholds and decisions are examples, not customer results or verified product demonstrations.

Three products, one lifecycle, one evidence trail

Assess it before it ships. Govern it while it runs. Keep the reason for every decision.

  1. Before deployment

    Trust Certify

    Should this agent ship?

    Discovery · Attack surface · Reliability · Reasoning · Scoring

    Our own assessment. Not third-party certification.

    Explore Trust Certify
  2. During execution

    Trust Runtime

    Should this action run?

    Validation, confidence and risk scoring on every proposed action, then a decision.

    Explore Trust Runtime
  3. Evidence and review

    Trust Audit

    What was decided, by which rule, and who acted?

    Append-only. Each record is hash-linked to the one before it.

    Explore Trust Audit
One evidence trail per agentPre-deployment results, every runtime decision, every reviewer and reason.

Works with the agent stack you already run

Add Fairway to an agent with one integration. Your data stays where it lives.

Integrations

One integration, three ways in: framework adapters, a zero-code auto-instrumenter, or a plain HTTP endpoint.

Read the integration docs

Deployment

Embedded in your process for development and CI, or self-hosted in your network with the full gateway and operator dashboard.

Read the deployment docs

Security and trust

Prompts, agent payloads and decisions stay inside your deployment. Every decision keeps its reason. Controls are mapped to the frameworks you answer to, as mapping, not certification.

Read the security docs

Find out what would have been held, before anyone asks you to decide

Self-serve pilot

For engineers

Run Fairway on your own agents with enforcement off. Nothing is stopped, you just see what would have been held.

  • EnforcementOff
  • PriceFree
  • RunsIn your own network
Baseline reportExample data
  • Refund a customer £4,800would have waited for reviewEscalate
  • Open a patient's recordwould have been stoppedBlock
  • Approve a £6,200 claimwould have gone aheadApprove

No sales call required.

Pre-deployment assessment

For teams

For teams that want Trust Certify run on an agent before it ships, walked through with your engineers, security and risk leads.

  • Discovery
  • Attack surface
  • Reliability
  • Reasoning
  • Scoring
AssessmentIllustration