The nine primitives

An agent harness decomposes into nine primitives. In regista each maps to exactly one module with one narrow interface — this table is the contributor's map: pick a primitive and you know which directory it lives in.

#	Primitive	Module	What it owns
1	Instructions	instructions.py	What the agent is told
2	Context management	context/	What the model remembers
3	Tools	tools/	What the model can ask for
4	Execution environment	environment/	Where effects happen
5	Durable state	trace/ + session.py	What survives a crash
6	Orchestration	loop.py	The turn engine
7	Subagents	— (v0.3)	Child agents with isolated context
8	Skills & procedures	— (v0.3)	Reusable instruction + tool bundles
9	Verification & observability	trace/ + policy/	Whether it behaved

The boundary rules

Three boundaries are subtle enough to state as rules:

Tools vs. environment. A tool defines what capability the model sees (name, schema, semantics); the environment defines where its effects happen (local filesystem today, a container tomorrow). Swapping environments never changes a tool's schema — the model can't tell the difference, and neither can a recorded trace.

Observability vs. durable state. One artifact — the trace — serves both. Replay, resume, and OTel export are all consumers of the same log; there is no second instrumentation path. If a behavior isn't in the trace, it's a bug.

Policy vs. tools. The permission gate sits in the loop, before dispatch. Tools never self-police; a tool implementation can assume it was allowed to run.

The dependency graph

Dependencies point strictly downward:

agent.py / session.py        (composition root)
        │
     loop.py                 (orchestration)
        │
providers/ tools/ policy/ context/ instructions.py
        │     └── environment/
     trace/                  (flight recorder + durable state)
        │
types.py + errors.py         (shared vocabulary — imports nothing)