02
Architect / Builder separation
One agent designing and implementing defends its own design. Splitting the roles keeps structure honest.
NOT WHEN · The structure is already fixed and the work is filling in a known shape.
WORKFLOWS
Agent Operating Patterns
A library of reusable patterns. Each defines the problem, the routing of work between human and agents, how the output is validated, and, as importantly, when not to use it.
PATTERN 01 · FULLY DEVELOPED
Research → Plan → Challenge → Implement → Review
RESEARCH→PLAN→CHALLENGE→IMPLEMENT→REVIEW
PROBLEM
Agents produce plausible code immediately, which makes it tempting to skip thinking. Unchallenged plans embed wrong assumptions that surface as expensive rework.
WORKFLOW
A research agent surveys the territory. A planning pass produces a stepwise plan. A separate agent, or the same one adversarially prompted, attacks the plan. Only a surviving plan reaches implementation. A cold reviewer reads the final diff.
HUMAN RESPONSIBILITY
Frame the problem, judge the challenge round, arbitrate disagreements, approve the plan, and own the merge. The human is the constant across every stage.
AGENT RESPONSIBILITY
Research breadth, plan drafting, adversarial critique, implementation within scope, and an independent cold-context review.
VALIDATION
The challenge round validates the plan; tests and independent review validate the code. Nothing merges on the implementer's word alone.
RISKS
Ceremony on trivial tasks; challenge rounds that degrade into agreement if the critic shares the planner's context; humans rubber-stamping surviving plans.
WHEN NOT TO USE
Small, reversible, well-understood changes. A one-line fix does not need a five-stage pipeline. That is what maintenance patterns are for.
03
Independent code-review agent
A reviewer with no memory of the implementation reads the diff cold and reports findings before any human review.
NOT WHEN · Generated throwaway code that will never reach production.
04
Test-first agent loop
Tests written and approved before implementation give the agent an objective target and the human a verification anchor.
NOT WHEN · Exploratory work where the desired behaviour is still being discovered.
05
Bug reproduction before repair
No fix is accepted until the failure is demonstrated, then demonstrated fixed. Kills "it should work now".
NOT WHEN · Truly non-deterministic failures. Then the pattern shifts to instrumentation first.
06
Documentation as executable context
Docs written so a fresh agent session can act on them: onboarding material that is tested, not assumed.
NOT WHEN · Fast-moving prototypes where docs would immediately rot.
07
Repository onboarding
A structured first session that maps the codebase, conventions and coupling before any agent is allowed to change it.
NOT WHEN · Repositories small enough to fit in a single context window.
08
Context compression & handover
Long tasks are split across sessions with deliberate handover documents, because context loss is silent and cumulative.
NOT WHEN · Tasks that comfortably finish inside one focused session.