Agent Lifecycle

Agent context has a lifecycle. Five layers, maintained in order.

The five layers

1. Registry — platform-neutral name, role, lane, status, reporting line, and pointers.
2. Soul file — stable identity, voice, boundaries, working style.
3. Memory — durable facts and learned preferences from prior work.
4. Skills — specific workflows or capabilities loaded only when relevant.
5. Runtime prompt — the current platform-specific expression of the role.

The runtime prompt is the most volatile. The registry is the most durable. Truth flows from registry down to runtime, not the other way.

Maintenance rule

Update the canonical source first. Then update the runtime expression.

If you correct a fact in a runtime prompt without updating the soul file or memory, the correction will not survive the next session. Worse, the next agent that loads from canonical sources will inherit the original error.

Why this design exists

HyperAgent, Claude, Codex, and the next runtime can all change. FIT keeps the agent operating model intact by maintaining the registry and source folders outside any single runtime.

A runtime swap should be a config change, not a rebuild.

Drift signals

Watch for these. Each is a signal that some layer is out of sync.

• A runtime answers from training rather than your soul file.
• Memory contains facts that contradict the registry.
• A skill loads on a task it does not apply to.
• The registry status says active but the soul file is empty.

When you spot one, fix the canonical source first. Then propagate.