The Harness is the architecture of the Seed Agents system: a local-first, account-scoped agent runtime that reduces to three kinds of state and five operations on them. These pages document it as built, including the parts that are designed but not yet built — those are marked as such.

New here? Read Architecture for the shape of the system, then The Five Verbs for its entire operation set. Looking up a term? Start at the Glossary.

Foundations

Architecture — components, principles, and the end-to-end path of a message

The Space — everything an agent has, as documents

The Log — everything that happened, actor-stamped

Runs — everything that executes, as a tree that is also the queue

Sessions and Threads — the conversation, and how it differs from a run

The Five Verbs — read, write, call, delegate, plan

The agent's world

Memory — persistent files under ~/memory

Tools as Documents — content-addressed tools, touch-expand, the Space index

Execution — one execute contract in hardware-isolated microVMs

Attachments — session-private files

Agents and Grants — definitions, permissions, signing identities

Orchestration

Delegation — model children with verbatim briefs, typed results

Scripts and the Journal — deterministic QuickJS and content-keyed replay

Plans — the visible checklist and runtime settlement

Obligations — what a run owes before it may end

Time and autonomy

Time and Parking — waiting is free: four wait reasons and their wake sources

SignalRun — answering a parked run by hand, exactly once

continueAsNew — day-scale loops that stay bounded

Triggers and the Event Bus — sources, continuations, exactly-once firing, loop guards

The platform

Model Providers — nine provider types, encrypted keys, OAuth, reasoning levels

The Signed Protocol — 52 actions, signed envelopes, the WebSocket

Persistence — the schema, migrations, and the parity test

Security — boundaries, injection defenses, and the real gaps

Hypermedia Integration — reading and publishing on the Seed network

The Desktop Surface — the run card, the transcript, the palette

Operating it

Operations — every environment variable, deployment shapes, troubleshooting

Testing and Gates — what must pass, and how missing coverage is marked

Limits and Constants — every bound with its constant name and value

Glossary — every term, defined once

The shape of the argument

Three nouns and five verbs is not branding. The old system was roughly twenty-five bespoke tools, each with its own prompt text and failure modes: every new capability made the prompt longer and the model's job harder, the user could only talk at the agent, and execution lived in memory where a restart erased it.

Under nouns × verbs, capability arrives as a document or an address, so the model-facing surface stays five verbs no matter how much the system grows. The user holds those same five verbs on the same log, so "the user did it" and "the agent did it" are one code path with different actor stamps. And because every execution is a durable row, waiting costs nothing and a crash is recoverable — which is what makes day-scale autonomous work ordinary rather than heroic.