intermediate · ai
Agent engineering: MCP, subagents and agentic systems
Agents as engineering rather than as demos: tools, context, memory, delegation, orchestration and evaluation.
·
By Amit Chakraborty — Kolkata, India · Remote worldwide
What you will be able to do
Design, build and evaluate a multi-agent system you would run unattended.
Syllabus — 14 chapters
- 1. What an agent is, minus the marketing — Reduce an agent to a loop, a model and a set of tools.
- 2. Tool definitions as a contract — Write tool schemas a model uses correctly on the first attempt.
- 3. The context window as a budget — Account for every token and decide what deserves the space.
- 4. MCP: what the protocol standardises — Understand what problem the Model Context Protocol solves.
- 5. Building an MCP server — Expose a real capability through the protocol.
- 6. Memory: persist, retrieve, forget — Give an agent continuity without giving it an unbounded prompt.
- 7. Retrieval that helps instead of distracting — Make retrieval improve answers, which it does not do by default.
- 8. Subagents: delegation, isolation and cost — Split work across agents and know what each split costs.
- 9. Multi-agent systems: useful, or theatre — Judge when multiple agents beat one good loop.
- 10. The orchestrator pattern — Coordinate agents with budgets, priorities and back-pressure.
- 11. Evaluation: proving the agent got better — Build an evaluation suite before tuning anything.
- 12. Failure modes: loops, invented tools, silent truncation — Recognise and instrument the ways agents fail quietly.
- 13. Permissions: what an agent may do unattended — Draw the line between autonomous action and human approval.
- 14. An agentic ecosystem end to end — Assemble the pieces into a system worth running unattended.
0 of 14 published. A new chapter every Sunday morning. Free to read.
Learn it free, elsewhere too
The documentation, videos, books and practice sites worth your time on this subject. All free. None of it is mine.