Agents
AI workers with durable runs, scoped context, tools and governed writes.
Primitive includes Agent Builder, durable runs, authorised context/memory, provider execution and a human approval path for agent writes.
An agent is like giving the AI a job, the right toolbox and a folder of information, then keeping a record of what it did instead of treating every answer as a disposable chat.
Agents turn AI from a response generator into an accountable worker. Their useful identity is the combination of role, context, tools, authority and execution history. Primitive separates the agent's reasoning/execution from the permission to make consequential changes.
Durability
Runs are durable and agent outcomes can be attached back to the originating communication context, giving the system an inspectable work history.
Authority
Agents use authorised context and tool boundaries. Human-approved writes and authority rechecks protect consequential changes.
Orchestration relationship
An agent may be used as a worker inside a larger orchestration. The recursive holon model extends this further by allowing a bounded worker/orchestrator structure, but that broader model is documented separately because it is not yet identical to the implemented Agent Builder.
Where this description comes from
These pointers are the authority behind the status and technical description. If implementation and documentation disagree, the canonical implementation or Control contract wins and this page should be corrected.