AI Agents in Speedboat

Use this guide if your team is using autonomous or semi-autonomous AI agents alongside Speedboat.

The short version: AI agents can accelerate work inside every lane, but they do not change who owns decisions or lower the finishing cost.

Speedboat helps AI-accelerated teams finish meaningful work, not just start more work. AI agents are useful when they help the team learn faster, make Run visible faster, and prove what Landed without weakening human judgement.

Core principle

AI agents are accelerators, not owners.

They can explore, draft, compare, generate, refactor, summarise, and automate parts of the work. They do not own product judgement, production readiness, integration risk, rollout decisions, or accountability for what lands.

Speedboat already assumes that starting work is cheap and finishing work is expensive. AI agents make that asymmetry stronger.

If agents generate more code, prototypes, fixes, docs, or upgrade proposals than the team can properly review and Land, the system slows down. The queue moves from “work to start” to “work that now needs judgement, integration, and review.”

The quality bar stays the same.

Preview

Preview is where agent autonomy is most naturally useful.

AI agents can:

generate quick prototypes or PoCs
explore multiple approaches in parallel
compare implementation options
produce demo-ready artifacts for discussion
summarise findings and trade-offs

Human ownership stays on the decision:

Should we build this?
What did we learn?
Do we kill, park, continue, or promote?

Treat agent-generated Preview output as disposable by default. Do not promote it into Build just because it appeared quickly. Promote it only when the team deliberately decides it is worth building properly.

Build

AI agents are most useful in Build when the task is bounded and the acceptance criteria are clear.

Good candidates include:

test generation and test expansion
migration scripts
repetitive refactors
boilerplate and scaffolding
documentation drafts
code search and comparison work

Human ownership stays on:

architecture and trade-offs
integration with the rest of the system
production readiness
review quality
rollout and support impact

The main failure mode is overproduction. If agents create more candidate changes than the team can review and integrate well, Build WIP rises in disguise. Treat review and integration as real work, not as free follow-up.

Run

AI agents can make Run work faster and more proactive.

Useful patterns include:

triaging alerts
summarising incidents
identifying likely root causes
drafting remediation options
proposing dependency upgrades
generating low-risk maintenance changes for review

Human ownership stays on:

deciding what action to take
deciding what gets deployed
deciding whether a fix is safe enough to roll out
deciding what follow-up work is needed

Do not let automation blur the deployment decision. Faster triage is good. Unowned rollout risk is not.

Guardrails

Keep agent-generated work inside the same WIP limits as any other work.
A WIP item is a meaningful unit of work the team is steering, not an individual ticket or subtask.
Prefer agents for bounded tasks with clear success criteria.
Count review, integration, and validation as real effort.
Do not treat agent output as production-ready by default.
Do not let Preview artifacts quietly become Build code.
Do not let Run automation bypass human deployment judgement.

Good questions to ask

“Is this reducing time to learning, or just generating more output?”
“Is this task bounded enough for an agent to help safely?”
“What review and integration work are we creating downstream?”
“Would we still trust this change if a human had to explain and support it tomorrow?”
“Are we speeding up finishing, or just speeding up starting?”

In one sentence

AI agents fit well inside Speedboat when they help the team learn faster, implement bounded work faster, and triage Run faster without weakening human ownership of quality, integration, and rollout.