agent task (AT)
definitionA precise, checkable spec an autonomous coding agent runs to do one bounded piece of work, halting rather than improvising when it meets a surprise.
The executable register: a surgical, command-verifiable spec an
autonomous coding agent reads to execute one bounded unit of work. Procedural (no
architectural decisions inside), executor-portable (names its model:), tripwired
(halts and surfaces on surprise), one commit per task. Authored via dandori. Child tasks
(Ho-NN-AT-MM.md) descend from a ho; standalone tasks stand alone.
The Two Registers. A ho document operates at the architectural register. It carries decisions, the reasoning behind them, deferred discoveries, and post-execution reflection. It is read by humans—the practitioner, future maintainers, anyone trying to understand why the system is the way it is. Its content is durable: it survives the project and gets revisited years later.
An agent task operates at the executable register. It carries exact files, exact changes, exact acceptance criteria, exact verification commands. It is read primarily by an autonomous coding agent—and secondarily by the practitioner reviewing the agent's output. Its content is operational: once the work is committed and verified, the task's job is done.
Mixing the registers in one document damages both. The architectural reasoning gets buried under schemas and signatures; the executable spec gets diluted by paragraphs of context the agent doesn't need. Extraction resolves the conflict structurally: each document speaks in one voice, and the relationship between them is preserved by explicit binding.