| Implement tool use and environment interaction | The biggest domain. Selecting and configuring agent tools and permissions. MCP servers in depth: adding a server as a tool, configuring a GitHub remote MCP server, MCP registries and allow lists. Scoping agents to repositories and branches, invoking agents from CI workflows, enabling autonomous branch and pull request creation, plus error handling, retries, rollbacks and escalation paths. | 20 to 25% |
| Prepare agent architecture and SDLC processes | Which steps agents should perform, common agent anti-patterns, defining inputs, outputs and success criteria. Separating planning from execution: structured plans, plan validation, and preventing action until a plan is checked and approved. Autonomy design and inspectable artifacts. | 15 to 20% |
| Perform evaluation, error analysis, and tuning | Expected outcomes and operational constraints, qualitative and quantitative evaluation signals, automated scanning for signals, failure analysis from logs, plans and traces, and classifying root causes: reasoning errors, tool misuse, or context and environment issues. Then tuning: revising instructions, refining memory and tool access. | 15 to 20% |
| Orchestrate multi-agent coordination | Orchestration patterns, agent isolation for parallel execution, and resolving conflicts: overlapping code changes, duplicated effort, contradictory outputs. Observability with logs and audit artifacts, handoff documentation, recovery patterns including rollback and human-in-the-loop, and the agent lifecycle: adding, replacing and retiring agents without breaking workflows. | 15 to 20% |
| Manage memory, state, and execution | Short-term versus long-term versus external memory, scoping memory to the task, expiration and pruning rules. Durable state artifacts, resuming work without repeating steps, and detecting context drift during long executions. Sharing state across tools while preventing conflicting or stale context. | 10 to 15% |
| Implement guardrails and accountability | Classifying agent actions by operational, security and compliance risk, assigning autonomy levels, identifying which actions need human judgment, blocking policy violations, least-privilege scoping, explicit authorization for irreversible changes, and preserving velocity by cutting approvals that do not reduce risk. | 10 to 15% |