M1 Field Notes · 7 min read
The four-part operating model for agent work
Why tasks, agent configurations, computers, and runtimes need separate responsibilities.
The coordination problem is not another prompt
A model can produce code, analysis, or plans, but a team still needs to know which request it served, who owns the outcome, what resources it used, and who accepted the result. Collapsing those questions into one chat makes work fast to start and difficult to govern.
Give each object one job
The task records the outcome and conversation. The agent configuration defines a repeatable way of working. The computer provides files, credentials, network, and tools. The runtime performs the actual model session. Separating them lets teams change execution infrastructure without losing the work record.
Review is a stage, not a sentiment
A completed run should return concrete evidence: what changed, which checks ran, what failed, and what remains uncertain. A named person can then accept, request changes, or stop. That decision is more useful than a vague claim that an agent is supervised.