miatzSign in

Splitting one job across several specialised agents with a coordinator

Rather than one long prompt doing everything, the work is cut into parts — one agent retrieves, one drafts, one checks the draft against the retrieved material — each with its own instructions and its own tools, and a coordinator passes results between them. Each part becomes separately testable; the handoffs between them become the system.

Effort
More than a few weeks
Skill level
Engineering skill required
Organisation size
Mid-market
Value
Quality, Time saved

Tools named for this

  • An orchestration framework with a coordinator and typed messages between roles
  • A single-model baseline kept running alongside, so the split has something to beat
  • Per-step traces retained long enough to tell which handoff dropped what

What to check before you ship it in India

  • Every hop copies the record into another context window, another queue and another log. Section 8(5) places the duty of reasonable security safeguards to prevent a personal data breach on the fiduciary, and a design whose whole point is passing material between components multiplies the number of places a copy comes to rest — each of which somebody has to be able to name.
  • The published taxonomy of multi-agent failures sorts fourteen modes into three groups, and two of the three — inter-agent misalignment and task verification — are about the seams rather than the models. A system that never verified whether the task was actually completed will report that it was.

Sources

Every claim on this page traces to one of these, on the date it was read.