Implementing an AI Agent: Five Stages from Task to Working Result
Stage 1. Investigation
The task, input data, expected artifact, acceptance criteria, constraints, and decision-maker are defined. The composition of the integrations, the environment and the cost are determined after the investigation — not before it.
Stage 2. Passport and permissions
The eight fields of the passport are documented: the task, the input, the process, the output, the limitations, the permissions, the environment and the verification. The sources, the allowed actions, the retention period and the human confirmation points are set. Unnecessary sources are not connected.
Stage 3. Installation and integration
Installation and configuration for the task, integration with the sources: email, the knowledge base, PostgreSQL, documents. For an agent this stage includes team training and configuring the access roles.
Stage 4. Launch with control
The first results pass a human check: the journal shows the inputs, the stages and the rationale for each conclusion, disputed decisions stop. The acceptance criteria from stage 1 determine readiness.
Stage 5. Support and refinement
An agent deployment includes 30 days of support. After that, changes to rules, sources, or formats are billed at $150/hour for a subagent and $300/hour for an agent. These rates cover services; infrastructure and subscriptions are billed separately. See the pricing page for details.
Adapted from the article: agentseffect.com.