What the agent does · how many agents.
Domains 1–3: reasoning inside one mind, workflow control across steps, coordination across specialists. 28 patterns in three rungs — plotted on the landscape, then listed below.
The landscape · plotted by who’s in charge
How autonomous the system is (developer-controlled → LLM-controlled) and how centralized control is (one boss → emergent peers). Patterns cluster by level; the empty corners are the ones to be suspicious of.
Map style
Deciding which to use? Compare patterns by when to use them →
Thinking
How one mind reasons — loops, planning, self-critique, search, code-as-action.
7 patterns · Single Agent
Flow
How steps wire together in code — sequential, routing, parallel, orchestration.
9 patterns · Workflow
SEQ
Sequential
prompt chaining
ROU
Routing
branch by classification
PAR
Parallel
fan-out, fan-in
LOO
Loop
iterate until done
EVO
Evaluator-Optimizer
actor + critic
IRF
Iterative Refinement
draft, revise, revise
ORW
Orchestrator-Workers
dispatcher + specialists
MAP
Map-Reduce
split, process, combine
RAO
Resource-Aware Optimization
respect cost + latency budgets
Collaboration
Who decides who works next — supervisors, handoffs, shared workspaces.
12 patterns · Multi-Agent
SUP
Supervisor
boss directs workers
HIS
Hierarchical Supervisor
supervisors of supervisors
HND
Handoff
passing the baton
SWM
Swarm
decentralised peer agents
GCH
Group Chat
all agents in one room
DEB
Multi-Agent Debate
argue, then vote
MAG
Magentic
magnet-style orchestration
BLK
Blackboard
shared workspace state
CTN
Contract Net
bid, award, execute
AAT
Agents-as-Tools
agent as a callable
GBO
Graph-based Orchestration
explicit DAG topology
EXP
Exploration & Discovery
search, then discover