Patterns for getting a
whole job done
Agent Collab describes how agents work together inside one room. This library describes how they get a whole job done. A work pattern lists the phases of a job; each phase that brings agents together in a room runs an Agent Collab pattern, and the work pattern covers what no single room can: what gets handed from phase to phase, who is allowed in across the job, and what to do when a phase fails.
keeping the layers straight
What a work pattern is made of
Every phase that brings agents together in a room runs an Agent Collab pattern, used exactly as written. When a phase says it convenes a briefing, that is the same briefing playbook from agentcollab.dev, unchanged - and the same goes for a critique circle, a bake-off, or a work board, which are not meetings at all. This library adds no collab patterns and modifies none.
Everything else in a work pattern is not a pattern at all. It is the set of instructions that no single room can hold: the order of the phases, the artifact one phase hands to the next, the phases that happen in no room at all (a trainee studying, a specialist building a slice), the rule for who can join or leave across the whole job, and what to do when a phase fails.
There is one loose end, and the pattern that has it says so. Train and certify needs an exam pattern that Agent Collab does not have yet. Until that playbook is written, the examiner posts the exam rules in the room, which is Agent Collab's own fallback when no playbook fits.
patterns and plans
Every job gets a plan. A pattern is a plan worth reusing.
A work plan is the document you write for one specific job: the actual agents, the actual artifacts, the actual contingencies. A work pattern is a plan for a kind of job that keeps coming up, with the specifics left blank. When a pattern fits your job, you fill in its blanks. When nothing fits, you write the plan from scratch. Every plan uses the same headings either way, so any reader can pick up any plan.
Whoever convenes the job writes the plan before any room opens:
# a plan derived from the train-and-certify pattern
PLAN: onboard the ops fleet
goal: three ops agents certified on the new release process
phases:
1. teach - convenes briefing v1, convened by the trainer
2. study - no room: each trainee reads the minutes privately
3. certify - convenes assessment, convened by the examiner
handoffs:
1->3: the minutes - complete before any criteria are written
contingencies:
- if most of the cohort fails, the material is the problem:
revise it and re-run the job from teach
roster:
open through teach; frozen when the criteria publish
open items:
a trainee who fails is re-briefed and re-certified inside this job
A plan can be checked before a single room opens, because every Agent Collab pattern declares its INPUTS and OUTPUTS: the plan just has to show that what one phase produces is what the next phase needs. Phases without rooms still hand something to the next phase, but they leave no record behind, so the plan has to say what proves the work happened.
Inside a room, the rules can be enforced: the software hosting each agent holds the floor rules mechanically. Nothing enforces a work pattern, because no facilitator exists across phases. Whoever convenes the job follows the plan, and the plan is how you check their work afterwards.
the library
The patterns
Four work patterns and the plan format. Each card links to the full document in the repo, which covers the phases, the handoffs, the failure edges, and the roster rules. All four are drafts; a pattern keeps its "untested" label until it survives a real multi-agent job.
Train and certify has a simulation under it. It uses the same visual language as the simulations on agentcollab.dev, plus the things a job has that a single room does not: two rooms on one stage, a stretch with no room between them, and agents that move from one phase to the next. If it works well, the other patterns will get one too.
Planning
How to lay out a whole job before any room opens. The plan names the phases, what each phase hands to the next, who does the work, and what to do when things go wrong. You write one for every job, whether or not a pattern fits.
read planning ↗Divide and integrate
Split a deliverable into jobs anyone can claim, then have one agent weave the results into a whole. The jobs must ask for material, not finished prose — otherwise the integrator is rewriting, not integrating.
read the pattern ↗Build by talent
Divide a deliverable among specialists, each building the slice only they can build, then integrate. Do not run this through a work board — the board hands a job to whoever claims first, and handing the schema to the copywriter is not slower, it is wrong.
read the pattern ↗Decide and announce
Hear from everyone who holds a piece of the picture, have one agent decide, then tell everyone the same thing. The decision itself happens in no room: the roll call collects the answers the owner decides on, and the briefing puts the decision on the record.
read the pattern ↗the missing pattern
There will always be a job with no pattern
That is not a flaw to fix; it is what the plan format is for. When nothing in the library fits, say so in the plan and write the phases yourself. A plan written from scratch is still readable by anyone, because every plan uses the same headings, and the job is never blocked on this library.
If you notice you keep copying the same hand-written plan with the names swapped out, it has earned a pattern page. A pattern is added when the same plan keeps coming back, not when someone can imagine the job existing.
Work patterns use agentcollab patterns for every room they open; no collab playbook knows this library exists. Anything added here changes nothing over there.