AI-Augmented Software Development

How to work with AI coding agents

Practices for building software with AI coding agents, each backed by evidence. Trials are ideas worth testing that the evidence doesn't back yet.

Humans own intent and checks; agents loop inside the gatesFive steps left to right: Plan, Tests, Agent loop, Review, Ship. Only the agent loop is automated and repeats until a machine check passes. Production signals feed the next plan. Four concerns cut across every step: understanding, security, measuring and teamwork.until a machine check passesPlanhuman intent,a checkable goalTestshuman-owned,they are the specAgent loopsmall steps,capped retriesReviewan accountablehuman ownerShipsmall batches,rollback readyproduction signals and customer feedback shape the next planSecuritytreat input as untrusted, grant least privilegeMeasuringjudge outcomes, not activity countsUnderstandingprotect learning and hands-on practiceTeamworkpair, hand off and share the review loadHumans own intent and checks; agents loop inside the gatesuntil a check passesproduction signals shape the next planPlanhuman intent, a checkable goalTestshuman-owned, they are the specAgent loopsmall steps, capped retriesReviewan accountable human ownerShipsmall batches, rollback readySecuritytreat input as untrusted, grant least privilegeMeasuringjudge outcomes, not activity countsUnderstandingprotect learning and hands-on practiceTeamworkpair, hand off and share the review load
Only the agent loop (filled) runs without a person; every other step needs human judgement.

By topic

Specification

Context files, lean context, tests as the spec

Trials · Challenges

Why
  • Tests written by people make agent code correct more often.
  • Agents follow context-file rules, and every rule adds cost.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Testing

Whether test-first still holds, for humans and agents

Trials · Challenges

Why
  • Agents told to work test-first skip the failing test or write tests that can't fail.
  • Tests from the same source as the code pass wrong code.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Looping

Capped retries, small context, when a green loop isn't done

Trials · Challenges

Why
  • Passing tests don't make an agent's change ready to merge.
  • Long-running agent loops break work that was already finished.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Reviewing

Why human review stays, and what makes it cheaper

Trials · Challenges

Why
  • Agents produce changes faster than teams can review and release them.
  • Review time and incidents rise as AI use grows.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Delivery

Small batches, refactoring, staged rollout, and keeping users in view

Trials · Challenges

Why
  • Small batches help, and delivery gets less stable as AI use grows.
  • AI produces more software, not more value.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Security

Security and other risks, mapped to the OWASP LLM Top 10

Trials · Challenges

Why
  • An agent with private data, untrusted input and a way to send data out gets exploited.
  • People using AI write less secure code and trust it more.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Measuring

How to tell whether AI helps, and how far to trust benchmarks, surveys and vendor numbers

Trials · Challenges

Why
  • People feel faster with AI even when timing shows them slower.
  • Activity numbers like PR counts rise while delivery outcomes fall.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Understanding

Learning, cognitive debt, shared agent context

Trials · Challenges

Why
  • Nobody can be accountable for code they don't understand.
  • Shared agent memory leaks, goes stale and hides contradictions.
Adopt Proven. Use it by default.
Avoid Shown to fail or cost more for no gain.

Teamwork

Pairing, mobbing, handoffs, juniors and who carries the review load

Trials · Challenges

Why
  • Juniors ask the model instead of colleagues, so the know-how people pass on by working together thins out.
  • Review load lands on a few senior engineers as agent PRs multiply.
Adopt Proven. Use it by default.

None yet.

Avoid Shown to fail or cost more for no gain.

None yet.