Structure over chaos. Convergence over chatter.
Cosmic turns one prompt into a small crew of specialists around one Leader. Turns are routed, memory is bounded, numbers are measured, and the run ends when the evidence says the answer is ready.
One run, start to finish
- in Your task optimised, and clarified if something important is open
- 1 call Planner profile, subtasks with dependencies, a crew of 2 to 5, tools
-
up to 30 turns
The conversation
- Select the next speaker: hand-off, request, dependency graph
- Speak: one agent answers, calling its tools first
- Record: statuses, decisions and measurements are tracked
- 1 call Finalizer reads the whole transcript, shaped by the kind of task
- out One answer with anything unresolved stated first
- Your words, exactly as you wrote them
- The Task Brief: objective, constraints, subtasks
- The ledger of decisions the crew has made
- The last 10 messages, shared by every agent
- Tool output and sources, stamped by the server
-
1
Plan the crew
One call profiles the task, breaks it into subtasks with real dependencies, and staffs the smallest crew that covers them: two to five specialists, plus one independent checker when correctness can be checked. Tools are handed to whoever needs them.
-
2
Pin the brief
The Leader's first message is the Task Brief D. It stays pinned outside the rolling memory, beside your own words and a ledger of decisions, so an agreement made on turn two is still binding on turn twenty.
-
3
Route every turn
The speaker selector is mostly arithmetic, not a model. A hand-off named in the last message wins; then an explicit request; then the most idle owner of a subtask whose dependencies are done. A model is asked only when none of those settles it.
-
4
Speak in a header the server can read
Every specialist opens with what it is addressing, its status, the decisions it made and what it needs from whom. The server tracks subtasks and the ledger from those lines, and the demo shows them as chips.
-
5
Measure, don't assert
Language models cannot count, and a model asked to check another tends to agree with it. So word counts, tallies and arithmetic come from tools, an armed agent must use its tool before it answers, and the server stamps the real output and the real source links onto the message.
-
6
Check in, then stop
Every five turns the Leader scores coverage C(t) and consistency H(t). The run ends when two evaluations agree the work is ready. If it stops improving instead, it closes anyway and the answer leads with what is unresolved.