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.
Do you like what you are reading? Subscribe to receive updates.
Unsubscribe anytime