Design and Implement the Mesh

Agent design should begin with the work and its accountability—not with a persona and a clever prompt.

Define the agent contract

For every agent, specify:

• mission and intended outcome.
• tasks inside and outside its boundary.
• lifecycle phases and triggering events.
• required skills and approved knowledge sources.
• tools and write permissions.
• inputs, outputs and handoffs.
• human owner and decision authority.
• evidence, controls and evaluation criteria.
• failure, escalation and shutdown behavior.

This contract becomes the basis for implementation, testing and governance.

Build shared foundations

A mesh needs more than separate retrieval indexes. It needs consistent identities for requirements, processes, decisions, solution components, controls, tests, risks and releases. The transformation knowledge model should retain relationships and provenance.

Shared services should provide identity, authorization, model access, tool execution, logging, evaluation and policy enforcement. This avoids every team inventing a different control pattern.

Make skills modular

Separate durable agent responsibilities from variable expertise. The Process / Fit-to-Standard Agent can combine general fit-to-standard capability with a Lead-to-Cash skill pack, a country localization pack and client-specific process knowledge.

Modularity makes capability reusable and evaluation more precise. It also prevents a proliferation of vaguely differentiated agents.

Design the workflow

Break the use case into observable steps. Mark where deterministic rules are sufficient, where model reasoning adds value and where a person must decide. Define recovery behavior for unavailable tools, conflicting evidence and low confidence.

Evaluate before expanding autonomy

Test with representative cases, difficult edge conditions and deliberately bad inputs. Measure factuality, completeness, traceability, policy compliance, consistency and usefulness—not only language quality.

Autonomy should increase only when both performance and operational controls justify it.