Skip to content
Agentic AI for serious engineers
Manifesto

The Serious Engineer's Manifesto

Ten principles for engineers who build agents that survive contact with reality.

  1. The default is not to use an agent.

    If a deterministic workflow gets you 95% there, that's the architecture. Reach for an agent only when the task requires reasoning the workflow cannot encode.

  2. Workflows beat agents on most measurable axes.

    Lower latency, lower cost, fewer failure modes, easier evaluation. The win condition for adopting an agent is dynamic judgment, not raw capability.

  3. Evaluation is the first artifact, not the last.

    Build the rubric and the test set before you build the agent. If you cannot describe what "good" looks like, you cannot ship one.

  4. A demo is not a system.

    Demos succeed in two minutes against one prompt. Systems succeed at 99th-percentile latency, malformed input, partial tool failures, and adversarial users.

  5. The trace is the truth.

    If you cannot inspect why the agent did what it did, you cannot debug, audit, or improve it. Tracing is not a nice-to-have.

  6. Cost compounds. So does latency.

    A 5-step agent on a 50-token tool budget at 2.5x retry rate is a $1 query that takes 18 seconds. Multiply by traffic. Decide if the value is there.

  7. Multi-agent is workflow with extra steps, unless coordination is the problem.

    Most multi-agent systems are sequential workflows pretending to be social. If the agents do not genuinely negotiate, you have not gained anything.

  8. Human-in-the-loop is architecture, not UI.

    Approval gates, escalation paths, audit trails — these are first-class system components, designed up front, not bolted on after legal complains.

  9. Memory is risk before it is feature.

    Cross-session state widens the blast radius of every failure: prompt injection, leakage, drift, replay. Treat memory like you treat a database table.

  10. Earn the complexity.

    Every framework, every tool, every model parameter you add must justify itself with measurable improvement. If the eval scores don't move, the complexity isn't worth it.

"Most agents are workflows in costume. Earn the complexity."

These ten principles are the spine of the book. Every chapter, Field Note, Recipe, and Lab Report tests one or more of them against real systems. When a principle conflicts with a finding, the principle gets updated and the conflict is documented.

Last revised: 2026-05-13.