Learning path
Agentic Engineering Fundamentals
Vendor-neutral loop: specify, plan, context, tools, execute, verify, retry, and approve.
- 1Task specificationA task specification defines outcome, scope, constraints, and verification for any coding agent—not just one vendor.
- 2Plan before executionHave the agent propose a short plan and file touch-list before it edits—so you can catch wrong approaches cheaply.
- 3Project instructionsProject instructions are durable repo-level guidance that coding agents load so they do not rediscover your process every session.
- 4Context pollutionContext pollution is when irrelevant files, logs, or tool output crowd the model’s window and degrade the next decision.
- 5Tools and resourcesIn MCP, tools are actions the model can invoke; resources are data the model can read.
- 6ExecutionExecution is the agent actually editing and running commands—bounded by the plan, permissions, and verification gate.
- 7Verification loopsA verification loop requires the agent to prove done with an external check—then fix—until the check passes or you stop it.
- 8RetriesRetries are bounded re-attempts after a failed verification—each attempt should use new evidence, not blind repetition.
- 9Human approvalHuman approval boundaries are checkpoints where an agent must stop for a person before risky or irreversible actions.