Stay in the loop
Explore, decide, and edit together when the problem is still taking shape
# coding agents
Three ways to put agents to work
The right shape depends on how much context, autonomy, and coordination the work needs
# choose the shape
Start with the smallest useful unit of agency
Hand over an outcome
Give one agent a bounded task with a clear finish line and verification
Split independent work
Run parallel agents only when their outputs can be owned and combined cleanly
01 / pair
Use conversation while the problem is fluid
Best for shaping intent, making trade-offs, and reviewing consequential edits
High context, low hand-off cost, immediate steering
pairing
YOUShow me the trade-off before changing it
AGENTOption A is simpler. Option B keeps the extension point
YOUTake B and keep the public API narrow
02 / delegate
Define the finish line, then get out of the way
Best for contained implementation, diagnosis, research, and maintenance work
01
Outcome
State what should be true when the work is finished
→
02
Boundaries
Name the files, systems, and decisions that are in scope
→
03
Evidence
Require tests or observable proof against the original symptom
03 / team
Parallelism pays only when ownership is clean
Best for genuinely independent research, review, or implementation lanes
One owner per output, explicit interfaces, one integrator
team
RESEARCHCompare the protocol options and cite constraints
BUILDOwn the server implementation and tests
REVIEWChallenge safety, UX, and failure handling
LEADIntegrate the contracts and resolve collisions
# the short version
Match autonomy to uncertainty
| Shape | Use when | Context | Default |
|---|---|---|---|
| Pair | The problem is changing | Shared live | START HERE |
| Delegate | The outcome is bounded | Packaged once | Scale this |
| Team | The lanes are independent | Split by owner | Earn this |