Use case

Let agents ship change.
Keep the approvals.

Coding agents open most of your changes now. Traffical gives them a governed path — declare, scaffold, propose, read the evidence — and keeps the calls that need a person with a person.

The problem

Your agents are fast, tireless, and have no memory of last quarter's incident.

An agent can open a pull request that changes a discount, a ranking weight, or a system prompt in minutes. What it can't do is know that pricing changes need sign-off, that refund_rate is the guardrail that matters, or that the same idea was tried in March and reverted.

So teams do one of two things. They let agents ship and discover the damage in a weekly review — or they gate everything behind a human, and the velocity they bought evaporates.

Neither is a governance model. Both are a missing layer.

without a control plane
# agent, Tuesday 02:14
git commit -m "tune discount 10 → 20%"
# no plan, no guardrail, no canary
# margin drops 6% for 3 days

# human, Friday standup
> "revenue looks weird, anyone know why?"
> "which change was that even?"
> "who approved it?"

# nobody. that's the problem.

What the agent does

One MCP server. The agent works the same surface a person does.

The agent has real capability: it declares parameters, wires them into surfaces, states intent, resolves a measurement plan from your certified protocols, and reads back the evidence at every phase. It writes the readout, flags the anomaly, and drafts the decision summary.

What it does not get is the ability to invent the metric or the threshold. Plans resolve from protocols your data team certified. If no certified protocol covers the change, traffic is blocked until a person approves the metrics and guardrails.

  • Intent is mandatory — the agent can't create a change without stating what it's for.
  • Risk is computed — from surfaces and bindings, not from what the agent claims.
  • Evidence is bound — a proposal carries the snapshot it decided on, plus an expiry.
  • Gates re-run at execution — approval never exposes traffic on stale evidence.
agent session — traffical mcp
// 1 — declare the parameter
traffical.parameters.create({
  key: 'checkout.discount_percent',
  type: 'number', default: 10,
  surfaces: ['checkout'],
})

// 2 — scaffold the change (intent required)
traffical.changes.create({
  intent: 'Increase checkout revenue by testing'
        + ' a more aggressive discount.',
  objective: { metric: 'revenue_per_session',
               direction: 'increase' },
})
// ← risk: medium (backend · decision binding)
// ← template: canary → experiment → rollout

// 3 — resolve the plan from certified protocols
traffical.plans.resolve({ changeId })
// ← Checkout Commercial v3, Pricing Safety v2
// ← 3 guardrails (2 blocking) · needs approval

// 4 — propose, don't execute
traffical.changes.propose({ transition: 'start' })
// ← proposal pending human approval

What the human owns

Fewer decisions, each of them consequential.

Certify the protocols

Which metrics must be measured, which guardrails apply per phase, what minimums gate advancement. Versioned and snapshotted onto every change that uses them.

Owner: Data

Pre-align the thresholds

Ship at X, stop at Y, discuss at Z — agreed before the change runs, so the readout doesn't become a negotiation.

Owner: Product + Data

Approve what propagates

Promoting a winner into the product default is always a human checkpoint. Strategy, taste, ethics, and close calls stay with people.

Owner: Change owner / Admin
Agent actionAutonomyWhy
Pause a phaseDirectSafety actions are never gated on risk class.
Revert a phaseDirectStopping harm is always cleared.
Reduce exposureDirectShrinking blast radius needs no approval.
Start trafficDirect if low risk + certified planThe only auto-approval cell. Anything else waits.
Expand exposureDirect on low risk onlyMedium and above become proposals.
Advance a phaseProposal by defaultNew changes default to propose.
Promote a winnerAlways a proposalWriting a product default is a mandatory human checkpoint.

One queue, whoever proposed it

Agent proposals, monitor recommendations, and teammate requests land in the same place.

Acme Storefront / Decisions / Pending 4 pending
Advance to rollout proposal
optimize-checkout-discount · agent · claude · evidence bound 14:22
↑ +4.2%
Promote winner to default human required
onboarding-step-reduction · agent · codex · 5 steps beats 7
↑ +6.8%
Approve measurement plan blocked
winback-subject-test · no certified protocol for Email surface
traffic blocked
Phase paused by guardrail auto-executed
ranking-relevance-tune · latency_p95 breached blocking threshold
contained

Every row carries the evidence it was decided on and writes a decision record when it resolves. The next agent in the seat reads that log before it proposes — which is how the platform stops repeating March's mistake.

Keep reading

The patterns behind this page.

Give your agents room to run

And your team the guardrails, evidence, and audit trail to stay confident while they do.