Roles and Specialization

A single generalist prompt juggling many concerns dilutes attention and conflates instructions. Decomposing a task into specialized agents — each with a narrow mandate, focused context, and tailored tools — sharpens behavior and isolates failure.

• Narrow scope -> more reliable, more steerable.
• Isolation -> a mistake in one role does not corrupt the others.

Specialization is divide-and-conquer applied to prompting.

An effective agent persona is more than a name. It declares: a mandate (what it owns), boundaries (what it must not do), inputs/outputs (its contract), and success criteria. Vague personas ('you are a helpful expert') under-constrain behavior.

Give each agent one reason to change its behavior. A 'research-and-summarize-and-fact-check' agent blends three objectives that pull in different directions. Split them so each can be tuned, evaluated, and debugged independently.

The test: if you cannot state an agent's job in one sentence without 'and', it is doing too much.

Specialization includes context isolation. A coder agent needs the code and the spec, not the marketing brief. Restricting each agent's context reduces distraction, cost, and the chance it acts on irrelevant material.

• Give only the inputs the mandate requires.
• Grant only the tools the role legitimately uses.

Least-context and least-tool are the multi-agent analog of least-privilege.

Roles should partition the work without overlap or gaps. Overlap wastes compute and creates conflicting outputs; gaps leave parts of the task unowned. Design the role set as a clean decomposition where every subtask has exactly one owner.

Draw the responsibility map first, then write personas to fill each cell.

Some of the most valuable specialists are adversarial: a Critic that finds flaws, a Red-Team that attacks the plan, a Verifier that demands evidence. Their mandate is explicitly to disagree, which counters the sycophancy and over-confidence of a lone agent.

Pair every generator with a skeptic whose success is measured by problems found.

Over a long session an agent can drift out of character, absorbing the tone or goals of others it reads. Reinforce identity: restate the mandate at each invocation, and keep each agent's prompt self-contained so it does not silently inherit another's instructions.

Treat the persona as something re-asserted every turn, not set once and forgotten.

Specialists must hand off cleanly. Define a typed contract for each role's output that the next role consumes. Loose prose handoffs cause misinterpretation; structured contracts make the pipeline composable and verifiable.

The Researcher emits cited claims; the Writer consumes exactly that shape; the Critic checks against it.

More agents means more calls. Specialization trades single-call efficiency for reliability and parallelism. Justify each role: if splitting does not measurably improve quality or enable parallel work, fold it back into a generalist.

• Parallelizable specialists can run concurrently -> latency win.
• Strictly sequential specialists add latency -> ensure the quality gain is worth it.

Too few roles and you have a muddled generalist; too many and you drown in coordination overhead. Pick the coarsest decomposition that still gives each agent a single, evaluable responsibility. Start coarse, split only when a role proves overloaded in evaluation.

Granularity is tuned empirically, guided by where errors actually cluster.

For each agent confirm: a one-sentence mandate, explicit boundaries, least-context and least-tool scoping, a typed output contract, complementary fit with no overlap or gap, per-turn identity reinforcement, and a justification that the split improves quality or parallelism. Adversarial critics earn their place by finding real flaws.

You notice one agent that researches, writes, and fact-checks is producing inconsistent, hard-to-debug results.

Decompose tasks into single-responsibility agents, each with a clear mandate, explicit boundaries, least-context and least-tool scoping, and a typed output contract. Partition the work cleanly with no overlap or gaps, pair generators with adversarial critics, reinforce persona identity each turn, and choose the coarsest granularity that keeps every role evaluable. Specialization buys reliability and parallelism at the cost of more calls — spend it where it measurably helps.