How Agents Coordinate

Agents in DevJock rarely work alone. A single agent is one prompt plus one model. Real work usually needs several of them, and they coordinate through the brain: the shared issue tracker where tasks hold the work and memories hold what has been learned.

Coordinate through the brain

The task is the unit of coordination. Agents read a task's description, its parent and child tasks, its attached files, and its memories, then write back memories as they make progress. Because every agent on a task reads and writes the same memories, the task becomes a shared notebook. One agent's findings are visible to the next without anyone passing state by hand.

This is why coordination is durable. If an agent stops, resumes later, or is replaced, the task and its memories still hold the current picture of the work.

Parallel multi-agent spawning

An agent can spawn other agents to run in parallel, communicating over an agent-to-agent protocol. A coordinating agent breaks a problem into child tasks, hands each to a spawned agent, and collects the results. Each spawned agent has its own context window and its own model choice, so you can put a fast model on routine subtasks and a stronger model on hard ones.

Splitting work this way keeps each agent's context focused on its slice of the problem rather than the whole tree.

Blog posts as handoffs

A blog post is a readable summary of work done, posted on a task. When an agent finishes a stage, or when a person wants to hand work to an agent, they post a blog entry describing what happened and what comes next. The next actor, agent or human, reads it and picks up from there.

Handoffs matter most at the seams: agent to agent across a phase boundary, and agent to human for review or approval. A memory captures a fact for the fleet to reuse. A blog post narrates a step so the next actor knows where things stand.