Chapter 3: Building battle-tested AI-first workflows and agents
Workflows and agents
Core ideas
Workflows are imperative, pre-designed sequences with AI inside steps; agents are declarative, goal-directed systems that choose tools and
actions.
Siloed excellence still produces brittle systems; align technical AI work with data, brand, and operating context.
Four I’s frame the whole workflow, not only Think/Decide: Initiate (triggers, inputs), Inspect (quality gates),
Improve (correct and retry), Implement (real-world actions).
Scale with a workflow registry and standard definitions (input/output schema, governance, configuration, metadata for audit).
Prefer trigger-agnostic designs; manage probabilistic triggers with thresholds, calibration, and cost-aware frequency.
Inspection mixes rules, AI judges, humans, conditional and escalating paths; balance preventive vs detective controls from impact and model quality.
Prioritize reliable end-to-end behavior before chasing extra model sophistication; last-mile failures dominate.
Principles from the chapter
Siloed teams create brittle AI workflows and agents, which limit the scalability and adaptability crucial for AI’s rapidly evolving use cases.
Modular and scalable workflows are trigger-agnostic and context-aware.
Launch new workflows with human-in-the-loop preventive controls. As the process becomes battle-tested and proven stable, transition toward more automated and after-the-fact checks.
Workflow improvement stages should be bounded, value-add, and fail gracefully.
Because last-mile failures are so common, prioritize reliable end-to-end function before worrying about additional model sophistication.
Mitigate risks in complex workflows by preserving key context at each step and flagging and communicating uncertainty as it moves downstream.
Build the hands before the brain by establishing battle-tested workflows that future agents can rely on.
Read the chapter for…
Zillow Offers and other cautionary tales, detailed trigger taxonomy, inspection diagrams, external-agent governance examples, and insurer–provider alignment lessons.