Examples

The examples/ module contains small, runnable harnesses. Each isolates one pattern; each has its own README; most accept -scripted and run a deterministic fake client so you can watch the machinery without an API key.

Example	What it demonstrates
quickstart	Canonical compiled source for the minimal Anthropic runner and typed weather tool used in the root README and getting-started guide. It needs a key and network only when run.
healthcheck	A world-verifying goal: the agent probes a fake server farm until every endpoint reports healthy. Schema + fan-out guardrails police the calls.
releasegate	Pre/post guardrails around a release process: the agent may only publish after every required check is green, and the goal confirms the publish actually happened.
hnupvote	Live browser automation under pursue: a real Chrome session where the oracle is verified world state and a login wall forces the re-drive path.
computer_use	Live observe → model → act loop using typed computer-use tools. Requires Chrome, network access, and an LLM.
spawn_worker	Hierarchical decomposition: a coordinator launches researcher / reviewer / coder tasks with agent_spawn, continues independently, and collects results explicitly under work-mode gates and a depth cap.
stuck_recovery	The decompose guardrail’s graduated response — pass, advisory, fatal — as an agent repeats a failing search, then recovers by delegating.
long_conversation	Compactor integration: a pressure-gated compactor keeps a long exploration inside the context window without orphaning tool calls.
shared_infra	No-LLM tour of code understanding, retrieval indexing/search, workflow graph execution, checkpoint saving, and a deterministic HITL review decision.
hitl_resume	Real checkpoint/review/resume boundary: save state, approve/deny/edit through HITL, reload, and invoke an explicit continuation.
local_mcp	Local stdio MCP subprocess: initialize, discover and call a tool, then disconnect cleanly.
dynamic_tools	Dynamic tool lifecycle: register from disk, invoke, reload, unregister, and reject a built-in name collision.
skill_catalog	Discover SKILL.md guides, load the versioned store, and select applicable skill content.
notify_drain	Live notification subscription plus deterministic offline drain after unsubscribe.
sandboxed_shell	Actual Linux Landlock enforcement: workspace writes succeed while an ungranted outside write is denied.
deterministic_trace	Runner lifecycle callbacks persisted as JSONL and read back without a provider.

Running them

# deterministic, no LLM
go run -C examples ./healthcheck -scripted

# shared infra tour, no LLM or external services
go run -C examples ./shared_infra

# live provider — flags are uniform across agent examples
go run -C examples ./healthcheck -provider anthropic -model claude-sonnet-4-6
go run -C examples ./releasegate -provider llamacpp

Provider selection goes through the same backends registry as everything else: -provider names a builtin, keys come from the standard env vars, -base-url (or LLAMACPP_BASE_URL etc.) overrides endpoints.

Reading order

Start with quickstart for the smallest compiled composition or shared_infra for the non-LLM infrastructure pieces, then healthcheck — it’s the smallest complete loop with a real goal. Then releasegate for guardrails as policy, and stuck_recovery for what graduated advisories look like in practice. hnupvote is the one to read when you want to see pursue driving something genuinely flaky (a browser) to a verified outcome.

Each example’s README walks through its harness construction; examples/patterns.md in the repo collects the recurring shapes — scripted clients, fake filesystems, oracle design — you’ll reuse in your own harnesses.