ACCORD — The control layer for autonomous intelligence

SECTION 01

[THE RACE]

Five labs are making intelligence autonomous

OpenAI, Google, Meta, xAI and Anthropic are racing to make intelligence autonomous. AI is turning into SI: systems that do not just answer prompts, but browse, code, use tools, reach into systems, spend money and keep working on their own.

Lab 01

San Francisco, CA

OPENAI

Buildingagents

Lab 02

Mountain View, CA

GOOGLE

Buildingagents

Lab 03

Menlo Park, CA

What can this agent access?

Answer: scopes

Question

02

What can it spend?

Answer: spend limits

Question

03

What actions can it execute?

Answer: action rules

Question

04

When does it need human approval?

Answer: approvals

Question

05

What exactly did it do?

Answer: the record

SECTION 03

[HOW IT WORKS]

Every action is checked before it happens

An agent asks to do something. Accord checks it against the mandate in milliseconds: scope, spend, action. Then it allows it, holds it for a human, or blocks it. Either way, the record is written.

01 · Requestprocure-07

02 · Mandate check0.0 ms

Scope—

Spend—

Action—

03 · Decision

ALLOW

04 · Recordsigned

SECTION 04

[TRY A MANDATE]

You set the rules. Watch the agents obey.

Move the limits. Flip the switches. The stream on the right is a procurement agent's day of requests, evaluated live against the mandate you just wrote.

Your mandateprocure-07

Per-transaction limit$5,000 Daily spend cap$20,000

Over-limit goes to a human

Allow new vendors

Allow production writes

Never: credentials, new agents

Live evaluation$0 spent today

Allowed

0

Held for a human

0

Blocked

0

SECTION 05

[THE PRODUCT]

Every agent gets a mandate.
Every action leaves a record.

Not another chatbot. Accord is Cloudflare + Stripe Dashboard + IAM for autonomous intelligence: it sits between your agents and everything they can touch.

Like Cloudflare

In front of every action

Every tool call, payment and system touch passes through Accord first, and is allowed, held or blocked before it runs.

Like Stripe Dashboard

Every dollar accounted for

Per-agent budgets, per-transaction caps, vendor allowlists and a live ledger of what each agent spent, and why.

Like IAM

Permissions by the verb

Read, write and execute rights per system and per dataset. Least privilege for intelligence that never sleeps.

Accord fleet overview console — Fleet overview · every agent, its mandate, its spend and every intercept

Mandate as codeprocure-07.mandate.yaml

Three outcomesnothing in between

Allow

Inside the mandate. It runs, and the record is written.

Hold

Over a limit. A named human approves or denies it, from a laptop or a phone.

Block

Outside the mandate. It never happens, and the attempt is recorded too.

Pause

One switch stops a single agent, a team of agents, or every agent in the company.

Accord approval queue — Approvals · a $18,400 wire held for a human

Accord mandate editor — Mandates · access, spend, actions, escalation

SECTION 06

[HUMAN IN THE LOOP]

Big decisions stay human

When an agent reaches past its mandate, Accord freezes the action and puts it in front of the right person, with everything the agent did to get there.

Accord approval on a phone — Approve from anywhere · Face ID

One held wireREQ-20261001-0418

9:37:40 AMAgent matches the invoiceprocure-07 runs a 3-way match on PO 88213. Allowed.

9:37:52 AMWire requested · $18,400Over the $5,000 per-transaction limit. Accord holds it. Nothing moves.

9:37:53 AMRouted to financeM. Okafor gets the request on her phone, with the agent's six steps attached.

9:41:07 AMApproved with Face IDNote to the record: "Bulk order for Q4 line."

9:41:07 AMWire released · record signedRequest, decision, approver and note are hash-chained into the audit trail.

Approvals today

Median wait

4m 12s

Decided by humans

212

Timed out · denied

3

If nobody answers30m → auto-deny + notify owner

Above $25,0002 approvers, one from Treasury

Approver channelsphone · Slack · console

SECTION 07

[THE RECORD]

What exactly did it do? Every answer, signed.

Every request, decision and approval becomes an event. Each event carries the hash of the one before it, so nobody, including the agent, can quietly rewrite history.

Audit chain3,820,114 events

Chain verifieded25519 signatures7-year retentionlog.delete: never

Accord audit trail — Audit trail · replay any agent session, export it for an auditor

SECTION 08

[THE ACCORD]

Four layers of controls and audits

The White House Accord on Super Intelligence explicitly calls for internal controls, monitoring, independent evaluation and oversight around frontier systems. Accord productizes that same control philosophy at the agent and action layer.

Internal controls to monitor models. An internal team to keep those controls operating. An independent external evaluator. A board committee that oversees them all. The four layers, summarized · Full text via Washington Examiner

The announcement@WhiteHouse · 30 Sep 2026

White House Accord on Super Intelligence
— The White House (@WhiteHouse) September 30, 2026

View the post on X ▸

Layer 01

Internal controls

Mandates enforced on every action. Agents cannot reach systems in ways they were never granted.

Accord: enforced

Layer 02

Internal control team

A console for the people who keep controls working: control tests, alerts and remediation.

Accord: operating

Layer 03

Independent evaluator

A read-only workspace for outside auditors, with signed exports and replay of any agent session.

Accord: read-only seat

Layer 04

Board committee

A quarterly oversight packet generated from the record, so directors see what agents did.

Accord: quarterly packet

Accord oversight view mapped to the four layers — Oversight · each layer produces its own evidence

SECTION 09

[THE STACK]

From the mandate to every layer of the AI economy

The White House sets the mandate. ACCORD is the coordination and intelligence layer that turns it into live controls across compute, models, cloud, cyber and government. Truth Signal feeds new policy back in as it happens.

The mandate

WHITE HOUSE SI MANDATE

Accord on Super Intelligence · four layers of controls and audits

Coordination · intelligence layer

ACCORD

Translating policy into mandates, monitors, audits and oversight

Compute

Training runs and inference capacity registered against the mandate before they start.

Clusters under mandate0

Models

Every frontier model version, its evaluations and any capability change before deployment.

Versions tracked0

Cloud

Agent runtimes, tenancy and network boundaries, enforced where the agents actually run.

Runtimes enforced0

Cyber

Controls around cybersecurity, threat monitoring and incident reporting, as the Accord asks.

Controls active0

Government

Board reports, independent audits and regulator-ready evidence, if the steps become law.

Reports filed0

TRUTH SIGNALLive policy and intelligence feed · via Truth API0 signals

SECTION 10

[WHO IT'S FOR]

Everyone who signs off on what agents do

The CFO

CAPITAL

Knows every dollar

Budgets per agent and per team
Caps per transaction
Vendor allowlists
A live spend ledger

The CISO

PERMISSION

Owns every scope

Least-privilege access
Egress and prod guards
Prompt-injection screens
A kill switch

The board

OVERSIGHT

Sees what happened

Quarterly packets
Incident escalations
Independent evaluator seat
Accord-mapped evidence

The builders

ACCOUNTABILITY

Ship agents faster

Mandates as code
Simulate against real history
Any model, any runtime
Approvals without a meeting

SECTION 11

[SECURITY ARCHITECTURE]

Defense in depth for autonomous intelligence

Accord is built on zero-trust principles. No agent is trusted by default, no agent holds standing credentials, and no action runs without passing six independent layers.

The stackoutside in

L1Agent identityshort-lived workload identity · mTLS

L2Isolationsandboxed runtime per agent

L3Egress controlper-agent network allowlist

L4Secrets brokerscoped, expiring tokens only

L5Signed mandatesversioned · two-person publish

L6Evidencehash-chained · externally anchored

Zero trust

Identity for every agent

Each agent gets its own cryptographic identity that expires in minutes. Every call is mutually authenticated.

No standing secrets

Agents never hold keys

A broker issues narrow, short-lived tokens per action. Nothing long-lived ever reaches the model's context.

Least privilege

Default deny

Anything not granted in the mandate is blocked, and every grant is scoped to a system, a verb and a limit.

Change control

Two keys to publish

Mandate changes are signed, diffed, simulated against history and need a second approver before they go live.

SECTION 12

[THREAT COVERAGE]

Every known agent risk, covered in layers

No single control is enough. Each risk is met by at least two of the four Accord layers, so a failure in one is caught by the next.

Coverage matrixthreat × layer · swipe →

Threat	01 Control	02 Monitor	03 Audit	04 Oversight
Prompt injection	Screen	Detect	Review
Tool misuse	Block	Detect	Review	Revoke
Data leakage	Egress	Detect	Review	Pause
Privilege escalation	Deny	Alert	Verify	Approve
Unvetted MCP servers	Registry	Verify	Review
Runaway spend	Caps	Alert		Approve
Model drift after upgrade		Detect	Re-test	Roll back
Log tampering		Verify	Anchor	Escalate

SECTION 13

[BUILT FOR THE SOC]

Agent events, in the tools your team already runs

Every Accord decision streams out as a structured security event, so your security operations center sees agents the same way it sees people and servers.

Event streamaccord.events · json

SIEM

SplunkDatadogElasticSentinel

Identity

OktaEntra IDSAML SSOSCIM

Telemetry + alerting

OpenTelemetryPagerDutySlackWebhooks

Data protection

In transitTLS 1.3 · mTLS

At restAES-256 · per-tenant keys

Key custodyHSM-backed · rotated

ResidencyUS regions by default

SECTION 14

[FRAMEWORKS]

One record, many auditors

Accord's evidence is organized against the frameworks your auditors already use, so one control produces proof for all of them.

Framework

White House Accord

4 layers

Framework

NIST AI RMF

Govern · Map · Measure · Manage

Standard

ISO/IEC 42001

AI management system

Guidance

OWASP LLM Top 10

application risks

Attestation

SOC 2

security · availability

Mappings show where Accord evidence supports each framework. They are not certifications.

SECTION 15

[ANY MODEL. ANY SYSTEM.]

Accord sits at the action layer, so it does not care which model is thinking. It governs what the agent touches.

SECTION 16

[QUESTIONS]

No. Accord does not think for your agents. It sits between them and your systems, and decides what each one is allowed to do, spend and touch.

A short, versioned file per agent: which systems it can read or write, how much it can spend, which actions need a human, and which are never allowed. You can simulate a new version against past activity before you publish it.

The Accord calls for internal controls, monitoring, independent evaluation and oversight around frontier systems. Accord puts that same four-layer philosophy into production at the level where agents take actions. Accord is an independent company and is not affiliated with the White House.

Any. Accord governs the actions, tool calls, payments and system access, so the model underneath can be from any lab or your own.

The mandate decides. The default is a timeout that denies the action and notifies the owner, so nothing slips through by waiting.

No. mandate.modify and log.delete are never-allowed actions, and the audit trail is hash-chained and signed, so any tampering is detectable.

SECTION 17

[THE THESIS]

Frontier labs are building intelligence. We are building the infrastructure that controls what intelligence is allowed to do.

▸ Launch app Try a mandate