F68 AI

Library · Chapter 04

Building an AI Agent

From a real job to an Agent that can be tested and improved.

Building an Agent does not start with choosing a model or writing a prompt.

It starts with a real job that needs to be done more reliably by AI.

The basic process:

Job → Input → Output → Workflow → Model → Knowledge → Tools → Test → Improve

  1. 6.1
    CHOOSE A JOB TO TURN INTO AN AGENTDon’t turn a job into an Agent if it happens only once and has no stable process.
    1 min read · Read →
  2. 6.2
    DEFINE THE OBJECTIVEWrite one sentence that clearly describes what the Agent exists to do.
    1 min read · Read →
  3. 6.3
    DEFINE THE INPUTThe clearer the Input, the easier the workflow is to design.
    1 min read · Read →
  4. 6.4
    DEFINE THE OUTPUTBefore writing the workflow, know what result the Agent must produce.
    1 min read · Read →
  5. 6.5
    DESIGN THE WORKFLOWA workflow describes how the job is carried out.
    1 min read · Read →
  6. 6.6
    SEPARATE MANDATORY AND OPTIONAL STEPSThis keeps the Agent from running unnecessary Tools or workflows.
    1 min read · Read →
  7. 6.7
    CHOOSE A MODELChoose the model based on the nature of the job.
    1 min read · Read →
  8. 6.8
    CHOOSE THE THINKING LEVELThinking depends on how hard the workflow is.
    1 min read · Read →
  9. 6.9
    PREPARE THE KNOWLEDGEProvide only the Knowledge the Agent truly needs.
    1 min read · Read →
  10. 6.10
    CHOOSE TOOLSTools are chosen according to the needs of the workflow.
    1 min read · Read →
  11. 6.11
    DESIGN HUMAN-IN-THE-LOOPNot every step should be left for the AI to decide on its own.
    1 min read · Read →
  12. 6.12
    DESIGN FAILURE BEHAVIORThe Agent needs to know when it must stop.
    1 min read · Read →
  13. 6.13
    TEST THE AGENTDon’t test an Agent with just one case.
    1 min read · Read →
  14. 6.14
    EVALUATE THE AGENT
    1 min read · Read →
  15. 6.15
    CREATE A TEST CASE SETSave the important cases.
    1 min read · Read →
  16. 6.16
    DEBUG THE AGENTWhen the Agent answers incorrectly, don’t fix everything at once.
    1 min read · Read →
  17. 6.17
    DON’T PATCH THE AGENT ENDLESSLYReview → Consolidate Rules → Remove duplicates → Refactor → Retest
    1 min read · Read →
  18. 6.18
    VERSIONING THE AGENTWhen the output changes, you can know exactly which component changed.
    1 min read · Read →
  19. 6.19
    DON’T CHANGE MULTIPLE VARIABLES AT ONCEFor example, an Agent is not performing well.
    1 min read · Read →
  20. 6.20
    WHEN SHOULD YOU SPLIT AN AGENT?
    1 min read · Read →
  21. 6.21
    MULTI-AGENTNot every problem needs Multi-Agent.
    1 min read · Read →
  22. 6.22
    DIFFERENT MODELS FOR DIFFERENT STEPSThe whole workflow isn’t required to use the same model.
    1 min read · Read →
  23. 6.23
    MEMORY AND STATEWhat the Agent knows over the long term.
    1 min read · Read →
  24. 6.24
    DON’T USE MEMORY TO STORE EVERYTHINGNot every conversation is worth remembering.
    1 min read · Read →
  25. 6.25
    FROM MANUAL TO AUTOMATIONDon’t automate an Agent from the very start.
    1 min read · Read →
  26. 6.26
    A SAMPLE WORKFLOWResearch a topic and produce a sourced report.
    1 min read · Read →
  27. 6.27
    AGENT-BUILDING CHECKLISTIf these parts are clear, the Agent has enough of a foundation to start running.
    1 min read · Read →