Curriculum

Agents and agentic architecture

Agents are 73 percent (68 companies) and enterprise AI agents are the single top product surface (11), so this is a central, mid-course chapter. The report asks for the anatomy a PM must reason about (tools and tool contracts, planning and

1. 06.01

   Workflow or agent

   Decide whether a problem needs an agent at all, and when it does not, pick one of the five workflow shapes that ships cheaper and more reliably.
2. 06.02

   The anatomy of an agent

   Name the five parts of an agent a PM must reason about, so you can locate a failure in the part that owns it instead of blaming the model.
3. 06.03

   The agent spec you own

   Write the agent specification a PM owns, naming what the agent may do without asking by pairing reversibility with measured reliability.
4. 06.04

   Write the tool contract

   The model never sees your code, only the tool contract, so the names, the allowed inputs, the capped output, and the error text are the real product you ship.
5. 06.05

   Inside the harness

   The harness is the code around the model, and fixing reliability there in the cheapest order, prompt then tools then harness, beats retraining the model almost every time.
6. 06.06

   Context as a budget

   Everything the model reads shares one finite window, so spend it on high-signal tokens and reach for retrieval only when the failure is missing information, not weak reasoning.
7. 06.07

   When multi-agent is worth it

   You justify a multi-agent architecture against its roughly fifteen times token bill, decide whether agentic search is a component in your agent or a product other agents call, and set the Task Time Horizon that defines a long-running feature's success, cost, and UX.
8. 06.08

   Reachable from someone else's agent

   You make your product reachable from a host you do not own by publishing an MCP server and, for agent-to-agent delegation, A2A, then decide deliberately whether it should learn from its own production data.
9. 06.09

   Structured outputs are a decoding constraint

   Structured outputs force the model's response to match your JSON schema at the token level, which guarantees the shape but never the correctness, and the schema you pick quietly changes answer quality and cost.
10. 06.10

    INTEGRATION CASE

    As a reviewer, you can check whether an agent memo names the unpredictable step, stop limits, and cost per successful task.