A Squad is a team of specialised agents that work the same problem together — sharing context over a blackboard, handing off in sequence, or coordinating under a manager agent. It’s how a real workflow gets done: not one model guessing, a team deciding.
Squads collaborate over a shared blackboard — a common working memory every agent can read and write — so specialists build on each other’s work instead of starting cold. Two coordination modes cover most real workflows: sequential pipelines, where agents hand off in order, and hierarchical manager-delegate, where a manager agent decomposes the goal and assigns it out.
Each agent keeps its own tools, guardrails, and human-in-the-loop gates — so a squad is more capable without being less governed.
Reason–act–observe agents with their own tools and guardrails, each good at one part of the job.
Deterministic steps that wire agents and integrations into a repeatable job.
Agents and workflows composed into a team that works together — shipped as portable JSON.
A common working memory every agent reads and writes — context is shared, not re-derived.
A manager agent decomposes the goal and assigns work to the right specialist.
Ordered hand-offs when the job is a chain — extract, decide, act, verify.
Each agent carries its own guardrails and configurable human sign-off on risky steps.
Every agent’s reasoning, tool calls, and decisions are logged — squad-wide, not per-agent silos.
A whole squad is portable JSON — export it, run it on-prem, or extend it.
A single agent does one job. A Squad is a team of specialised agents collaborating over a shared blackboard — sequentially or under a manager agent — so a whole workflow gets done, not one step. For single-agent patterns, see the AI Agents page.