Agent Role Contract Definition (Onboarding)

Agent Role Contract Definition (Onboarding)

1. 01

   Map the client workflow into discrete task units before naming any agent

   List every handoff currently performed by a person: data pull, draft, review, publish. Each handoff becomes a candidate agent boundary, and the count tells you whether orchestration is justified at all.

2. 02

   Write a one-page role contract per agent covering inputs, outputs, and refusal conditions

   Specify the exact schema the agent accepts and returns, plus the cases where it must stop and escalate. Raft's persistent-agent model shows why memory scope belongs in this contract: an agent that retains codebase or client context needs an explicit boundary on what it may recall.

3. 03

   Assign a named human owner to each agent, not to the workflow as a whole

   When a chain breaks, the first question from a client is who was responsible. Per-agent ownership shortens that conversation from days to minutes.

4. 04

   Classify each agent by autonomy level and gate the high-risk ones

   Any agent touching client-facing communications, CRM records, or spend should require a human checkpoint before execution. Document the classification so it survives staff turnover on the retainer.

5. 05

   Define the fallback path for every agent before the first production run

   Decide in advance whether a failed step retries, degrades to a simpler model, or halts the chain and pages the owner. AgentX-style evaluation pipelines exist precisely so this logic is tested against a fixture set rather than discovered live.

6. 06

   Confirm where client data is processed and record it in the account file

   Self-hosted options such as Pacific Slate keep retrieval inside infrastructure the agency controls, while hosted platforms route through vendor tenancy. Either is defensible, but the client contract should state which one applies.

7. 07

   Run the full chain end to end on a synthetic brief and time it

   Capture baseline duration and correction count. The 40 to 60 percent timeline reduction that makes orchestration sellable is only credible once you have measured it on one real workflow.