Agents & Cognition · 06.11
Swarms & Converge
Fan a task out to worker sessions, each in its own worktree, then merge — the two cooperation shapes and when to reach for each.
Two shapes for two different problems
Swarms (agent_swarm, dynamic_swarm) fan a task out to independent worker sessions, each in its own git worktree, then merge their results — right when several genuinely parallel attempts at the same problem are worth running at once. Converge (and its smarter sibling pursue) instead iterates a single cohort sub-graph repeatedly until a continuation policy stops it, sharing one blackboard and one worktree across every iteration — right when the work is inherently sequential refinement, not parallel exploration. The full mechanics of the convergence loop itself — stagnation detection, ASSESS, REFLECT — live in Reliable by Design; this page is about when each cooperation shape is the right tool.
Continue
How a subdomain of work gets deterministically mapped to the specialist accountable for it, with every hand-off recorded and auditable.
How an agent hands a bounded task to a resolved teammate and awaits the result, with budget and attribution threaded through so no spend is anonymous.