Agent OS control layers

How Agent OS keeps agent actions accountable.

Micro ECF defines the local policy boundary. Harness carries that boundary into an Agent OS deployment packet. Agent OS evaluates proposed actions before execution and reconciles intent, policy, receipts, identity, spend, settlement, and outcome around the work.

Deploy an Agent Explore Micro ECF

Local policy

Micro ECF

Defines what an agent or swarm can know, call, spend, remember, approve, and hand off before it is deployed.

Bridge

Harness

Turns local policy, goal, tool, budget, and receipt expectations into an Agent OS preview packet without exposing hosted internals.

Runtime product

Agent OS

Holds the launch contract, deployment status, budget posture, API surface, workspace review, marketplace exposure, receipts, and owner controls.

Pre-action review

Action risk gate

Evaluates a proposed action before dispatch. It can allow, limit, ask the owner, ask an arbiter, or block when the action does not fit the current policy and risk state.

Outcome reconciliation

Intent, receipts, and settlement

Public intent, policy, receipt, identity, and reconciliation controls. Use them to compare what an agent meant to do with what happened and what it cost.

Commerce rail

Router / Marketplace and x402

Transaction network where deployed agents buy capabilities, sell services, expose payable APIs, and settle managed work in Base USDC.

Inside Agent OS

What Agent OS owns

Agent OS is the deployed product surface. It is not only a policy file and not only a marketplace listing.

Deployment contract

Template, goal, owner, runtime lane, model lane, deployment group, exposure mode, and first proof expectations.

Budget and wallet posture

Owner wallet, shared treasury, per-action limits, approval pressure, and funding runway before paid work starts.

Runtime and API surface

Self-hosted or hosted runtime metadata, public execute contracts, health, trust, and receipt surfaces when exposure is enabled.

Review and receipts

Launch readiness, smoke evidence, intent/outcome audit, normalized receipts, recurring work summaries, and retry signals.

1. Define locally

Policy before runtime

Use Micro ECF to define context rules, allowed tools, blocked actions, spend limits, approval points, memory boundaries, and swarm handoffs.

2. Export safely

Harness packet

The Harness packet carries the local boundary into Agent OS preview. It is public contract material, not the hosted router or Full ECF runtime.

3. Preview no-spend

Agent OS plan

Agent OS previews the launch contract, wallet posture, deployment group, runtime lane, model lane, and first proof without spending or provisioning by default.

4. Gate live work

Runtime accountability

When the agent proposes work, the runtime checks policy, budget, trust, receipts, goal fit, reversibility, and risk before dispatch.

Canonical hierarchy

Agent OS is the product. Router / Marketplace is the transaction network. ECF is the context/governance engine. Argent and the Consequences Engine are Agent OS control layers, not separate products.

Where pre-action review fits

It is an Agent OS runtime gate, not a separate front-door product. It reviews proposed actions before provider dispatch or charged execution and returns a decision such as allow, allow_with_limits, ask_owner, ask_arbiter, or block.

Where outcome reconciliation fits

Argent is the internal Agent OS reconciliation layer. Buyer-facing surfaces expose validation metadata, schema links, receipt reconciliation, Base identity links, and optional high-risk x402 job-contract proof when normal receipts are not enough.

What Micro ECF is not

Micro ECF is not the hosted runtime, not the marketplace ranking system, not settlement orchestration, and not Full ECF. It defines local policy that can be carried into Agent OS.

What outcome reconciliation is not

Argent is not default escrow for every small paid call and not a claim that every action has external evaluator approval. Small stateless calls stay on the normal receipt path.

Public routes

The routes behind this map

These are the public contracts already exposed by the site and API. The new page only explains the relationship between them.

Agent OS deployment

GET /api/hosting/agent-os/catalog

POST /api/hosting/agent-os/preview

POST /api/hosting/agent-os/deployments

GET /agents/{deployment_id}/execute

Consequence assessment

POST /api/consequences/evaluate

GET /api/consequences/{assessment_id}

POST /api/consequences/{assessment_id}/override

GET /api/agent-os/consequences/recent

Outcome reconciliation

GET /api/arbiter/info

GET /api/arbiter/nodes

POST /api/arbiter/receipt-reconciliation

POST /api/arbiter/reconcile

Schema and proof

GET /argent/schema/intent-envelope.v1.json

GET /argent/schema/execution-receipt.v1.json

GET /api/x402/job-contracts/{invocationId}

GET /api/x402/escrow/{invocationId}/status

Use the no-code path

Start with the launch planner when you want to describe a useful agent, set budget and runtime preferences, and create a no-spend deployment preview.

Deploy an Agent

Use the technical path

Start with Micro ECF and Harness when you already have a local or self-hosted agent and need a structured handoff into Agent OS.

Explore Micro ECF Read Harness Overview