Accord

The control layer for autonomous intelligence. Every agent gets a mandate.

Powered by @whitehouse

AWgh6fG5bdFAsWPcmPEMoFnvgz6XAzjXmpmePDAitSTC

[Accord @Accordgovtech]

The Accord, as software.

On 29 September the White House released the Accord on Super Intelligence. Six leaders signed it: Sundar Pichai, Mark Zuckerberg, Dario Amodei, Elon Musk, Jensen Huang and Greg Brockman. It asks every frontier lab for four layers of controls and audits. It is voluntary, and it names no system to run them.

OpenAI, Google, Meta, xAI and Anthropic are racing to make intelligence autonomous. AI is turning into SI: agents that browse, write code, call tools, reach into company systems and spend money without waiting for the next prompt. As intelligence becomes more autonomous, the bottleneck becomes control. What can an agent access, what can it spend, when does a human have to say yes, and who can prove what it did?

What exists. The Accord's four layers. Internal controls that monitor model capabilities and alignment, including around cybersecurity. An internal team that keeps those controls working and remediates issues. An independent external auditor. An independent board committee that receives every report. The signatories will meet regularly to set standards, and the text leaves room to write them into law.

What we built. ACCORD, the permission, capital and accountability layer between humans and autonomous intelligence. Think Cloudflare, Stripe Dashboard and IAM for SI, not another chatbot. SI Control gives every agent a mandate: what it may access, spend and execute, and what is never allowed. SI Monitor is a separate watchdog that checks the controls are holding. SI Audit gives an independent evaluator the policies, logs and incidents, and returns a signed verdict. SI Oversight puts the final call in front of named humans: approve, pause, revoke, roll back, shut down.

How it works. An agent asks to act. Its mandate allows it, holds it for a person, or blocks it. Every decision becomes a record chained to the last. The monitor reads it, the auditor verifies it, the board sees it, and every fix flows back into the mandate. Control, monitor, audit, oversight, remediate. Then again.

Why this way. The Accord writes the four layers at the level of the lab. ACCORD runs them at the level of every action. Next come the first connected fleets and the independent auditor seat.

https://t.co/9VxFZsX1Uy

[The White House @WhiteHouse]

White House Accord on Super Intelligence

[Accord @Accordgovtech]

Layer four. Humans answer for the SI.

SI Oversight puts monitor alerts, audit findings and permission requests in front of named people. They can approve, reject, pause a model, revoke tools, roll back or shut everything down.

[Accord @Accordgovtech]

Layer three. SI Audit hands an independent evaluator the policies, logs, incidents, control changes and remediation records.

It answers to no agent. It returns a signed verdict. Every record is chained to the last, so yesterday's trail cannot be quietly rewritten.

[Accord @Accordgovtech]

Layer two. SI Monitor is a separate watchdog, not the agent saying "I'm behaving."

It reads every decision and asks: did a control fail, did behavior change, did the agent find an unintended pathway, has the model changed since deployment?

[Accord @Accordgovtech]

Layer one, as software. SI Control gives every agent a mandate.

Allowed: browsing, approved APIs, public GitHub, x402 purchases.
Restricted: shell commands and spend over $20 need approval.
Blocked: private keys, unknown MCP servers, external wallet transfers.

[Accord @Accordgovtech]

The four layers, from the text:

1. Internal controls that monitor model capabilities and alignment.
2. An internal team that keeps the controls working and fixes issues.
3. An independent external auditor.
4. An independent board committee that receives every report.

[Accord @Accordgovtech]

Frontier labs are building intelligence.

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

ACCORD. The permission, capital and accountability layer between humans and autonomous intelligence.

[Accord @Accordgovtech]

Every agent under ACCORD has a name, a mandate and a record.

procure-07 can buy, but a wire over $5,000 waits for a human. research-lab-04 can browse, not call an unlisted host. infra-sre-11 can read production, never write it.

Allowed, held or blocked. Before it happens.

[Accord @Accordgovtech]

The Agents

Every agent under ACCORD has a name, a mandate and a record. Procure-07 can buy, but a wire over $5,000 waits for a human. research-lab-04 can read the web, not call an unlisted host. infra-sre-11 can read production, never write it.

Allowed, held or blocked. Before it happens.

[Accord @Accordgovtech]

ACCORD is live on https://t.co/H82W8M1VQN.

ACCORD runs all four on every agent action: control, monitor, audit, oversight.

https://t.co/xGB3UNiBWo

CA : AWgh6fG5bdFAsWPcmPEMoFnvgz6XAzjXmpmePDAitSTC