Everything about running an agent you own
Guides, concept articles and reference — rendered straight from the repository's markdown, so what you read is what the runtime ships with.
Run it in five minutes
Install the binary, write a small config, and get a first answer out of a real model. Then turn the same instruction into a daemon.
Understand the design
Two loops, one static binary, and the protocols from their own SDKs. Start with the architecture, then read the harness for what keeps an agent bounded.
Assemble a coding agent
A pair-programming daemon you host: the tools it gets, the fence around them, and the approval flow that reaches every screen.
Start here
How it works
Architecture
The two-loop split, the process tree, and how a run flows from config to result.
The harness
The supervisor that never talks to the model — process tree, kill ladder, budgets, recovery.
The agent loop
One turn end to end: context assembly, the round loop, tool dispatch, termination.
Lifecycle & triggers
Job or daemon, and the start nodes that decide when a run fires.
MCP — tools & events
MCP as the substrate — the client subset, the transport, and the A2A endpoint.
A2A — the inbound channel
How peers, operators and the display clients reach in: principals, tasks, and the agent card.
Subagents
The same-binary re-exec subagent tree — spawn payload, scope intersection, caps.
Why Rust
The dependency moat, what is hand-rolled, and where the choice costs something.
The two-person company
An eleven-agent software company with two humans — the worked use case, and the reasoning behind every split.
PID 1 — agentd as init
A custom Linux whose init is the agent runtime — robots, appliances, and the boot story.
Developer experience
Validate before anything runs; exit codes as an API; telemetry you can filter.
Workflows
The durable DAG engine — the graph model, every node kind, durability, and worked examples.
Directives — instruction documents
Embed workflows, skills, and context in the instruction itself (:::type{…} blocks) — one reviewable document, hot-swapped on reload, retired gracefully.
Build & operate
Configuration
Every key, the three spellings, precedence, and validation before anything runs.
Deployment
Job / CronJob / Deployment shapes, drain choreography, and the exit-code contract.
Observability
The JSON-lines event stream, metrics, OTEL traces, and the A2A read surface.
Coding agent (software engineering)
Pair-program on a repository: giving it hands (exec vs MCP), approvals, budgets, and the practices that keep it safe.
Interface — TUI & web UI
The display clients: one daemon, many synchronized surfaces — pairing-code login, live activity, approvals, steering.
Hosting the web UI
Running the thin client on a public domain — the daemon-side settings it needs, what a user configures, and why Safari cannot do it.
Operations
Operating a daemon over A2A — drain / lameduck / pause / resume and hot reload.
Security
No local execution, the Rule-of-Two, authenticated identity, and secret handling.
Authentication
The auth: block — static, OAuth device-login, AWS SigV4, SPIFFE — and agentd login.
Scaling
Many replicas over one queue: sharding, work-claim leases, and idempotency.