Four structural rules that allow parallel agents to split, work independently, and merge back into a single coherent result.
Adapted from @monokernI PUT GPT-6-ASTRA INSIDE 24 AGENTS. GETTING THEM TO WORK WAS EASY. GETTING THEIR WORK BACK INTO ONE DECISION TOOK THE SYSTEM. one objective enters as a single mass system finds the parts that can move independently, separates them, and assigns each fragment its own context, tools and completion condition agents spread out, work in parallel and return with artifacts instead of chat messages then Astra has to recombine everything this is where most multi-agent systems quietly fail outputs can all be individually correct and still refuse to become one coherent result. two agents make incompatible assumptions. three solve the same problem. another finishes perfectly after its decision has already become irrelevant so I added four rules: - tasks only split when they can proceed without waiting on each other - every result returns with evidence, dependencies and unresolved questions - the synthesis agent can reject an output and send it back for repair - temporary agents disappear after the merge instead of remaining in the system forever that last rule changed the behavior more than adding another model the team stopped accumulating roles and started forming around the shape of each objective sometimes Astra becomes three agents sometimes twenty-four sometimes the correct move is to stay as one the liquid structure is the actual execution state: fragments separating during delegation, moving independently during work, then colliding and merging as the final answer is assembled my take: agent count is a vanity metric until the system knows when to divide and how to become whole again grok bot blueprint that led to this astra experiment is in the article below
