AI Assisted Software Engineering

Part IV · Practice

18 Running the Whole Life Cycle: The Engineer as Manager of an Agent Team

Part IVLab: Three claims, your dataFeeds FinalP1P2P4P5P8

Learning objectives

After this chapter you should be able to:

  1. Run a semester-length project as the manager of a team of agents, and state for each milestone what is delivered and what a passing piece shows.
  2. Keep a work log from which cost-model parameters, defect data and handoff measures can be computed.
  3. Write a handoff package another engineer can work from, and take over a classmate's project by establishing an as-received baseline.
  4. Score pre-registered estimates against actuals, separating fitting after the fact from predicting a project the model has not seen.
  5. Write a blameless post-mortem and defend your decisions orally, citing your own data.
  6. Use the thirteen verdicts to test three claims about what limits software engineering in the agent era, and name the signposts that would change each.

Opening case

A composite case, not a real engagement.

Tomas builds an equipment-loan tracker for a campus makerspace: about 70 function points by his own count, roughly 3,500 lines of Java. Before any construction he commits four estimates of his total human hours. COCOMO II at nominal settings says about 1,770. With the tool multiplier at very high, 1,380. Story points at the pace he measured in his first week, 68. The agent-aware model from Chapter 6, with its illustrative parameters and about 18 agent tasks, 63.

The chapter continues from here.

What the chapter covers

  1. 18.1 The engineer as manager of an agent team
  2. 18.2 The project and its milestones
  3. 18.3 The work log
  4. 18.4 The handoff and the inheritance
    • 18.4.1 The rotation and the package
    • 18.4.2 Questions through a logged channel
    • 18.4.3 The as-received baseline
    • 18.4.4 The handoff audit
    • 18.4.5 What the handoff measures
  5. 18.5 The post-mortem and oral defense
    • 18.5.1 The blameless post-mortem
    • 18.5.2 The oral defense
  6. 18.6 The reveal
  7. 18.7 What changed and what didn't
  8. 18.8 Teaching and working for a moving target

Lab and prediction

Lab: Three claims, your data

Feeds milestone Final (uses M6 data). How the labs fit the project

The lab tests these predictions from the working paper the book grows from (Smith 2026a):

  • P1 Share of human hours on review and verification exceeds the share on construction from the first milestone on
  • P2 Defects traced to missing or ambiguous specification are the largest root-cause category
  • P4 Agent-aware cost model has lower error against actuals than COCOMO II calibrated on delivered LOC; tested on held-out projects against COCOMO II at its strongest
  • P5 Earned value at code-complete and at verified merge diverge, most in weeks of highest agent output
  • P8 Fuller (more complete, not longer) convention files and decision records at handoff predict lower comprehension hours for the inheritor

A result that contradicts a prediction, honestly reported, is a result. You are graded on accurate records and reasoned interpretation, not on agreeing with the book.

Summary

  • You are graded as the engineering manager of a team of agents, on specification, decisions, evidence and records, not code volume or agreement with the book. A log that contradicts a prediction is a result.
  • The work log is the dataset for the cost model, the growth-model fit, the handoff measures and the synthesis. It is worth keeping only if timely, honest and never edited.
  • Inherited work is graded against the as-received baseline. Score packages on accuracy, specificity and usefulness, not only completeness, and treat classroom correlations as exploratory.
  • An actual inside a refitted range is compatible with the model, not proof of it. Test models by predicting held-out projects, with every model given its best calibration.
  • The thirteen verdicts summarize the audits; their value lies in their assumptions and tests. Three claims (the loop, decisions, coordination) can be tested on your log, and each verdict has a signpost that would change it.

Key terms

Engineering manager of an agent team
your role: you specify, decompose, assign, review, verify, decide, escalate and record, and answer for the result.
Handoff package
the specification and backlog, convention file, decision records, risk register, work log and open issues that travel with a project, plus the baseline record and characterization-test pointers.Also used in 10, 13, 14
As-received baseline
the tagged state of an inherited project plus the characterization tests recording what it does.Also used in 11
Handoff rubric
accuracy, specificity and usefulness, each scored 0 to 2 per package item.
The reveal
the scoring of pre-registered estimates against actuals at the end of the term.Also used in 6
Held-out test
predicting a project the model was not fitted to; a refit that covers the actual shows only compatibility.Also used in 6

Connections

Where this chapter and another point to each other most often.

See it on the map

Further reading

  • Hassan (2026), §11.4 (p. 333) and the Afterword (pp. 347 to 349). The decision-bottleneck claim and the human limits of managing a fleet.
  • Tantithamthavorn (2026), §6.9 and §10.9. The loop, and the post-mortem format.
  • Feathers (2004), Chapter 13. The origin of characterization tests, and a guide to changing code you do not understand.
  • Dewar (2002). How to keep an assumption audit alive after the first pass, by choosing signposts and watching them.