CoralAI Labs
Earlier coordination study · September 2026
Research agents can repeat work or miss another agent’s findings. This study tested whether sharing and reviewing results helped them improve a program within a fixed testing budget.
Follow a GPU optimisation experiment
The goal: perform the same calculation using fewer GPU cycles.
Explore
Four Claude Code agents work on a faster GPU kernel—the code that tells a GPU how to perform a calculation.
Each agent proposes experiments.
Four agents investigate candidate approaches in parallel.
Test
An experiment produces evidence the agents can use to choose what to try next.
234evaluations per configuration
The budget stays the same in both runs.
Each candidate is evaluated. The measured result returns to its agent.
Compare
In this run, the agents met every 16 evaluations to share evidence and agree on different directions.
A shared blackboard becomes a working discussion.
Evidence moves from all four agents into a shared meeting and back to the team.
Improve
Agents return to their experiments with the team’s findings, not just their own.
Separate investigations. Shared learning.
The agents pursue distinct next experiments after exchanging findings.
The result
A faster kernel from the same four agents and the same evaluation budget.
1,159 → 1,097 GPU cycles.
We changed how the agents worked together.
The experiment
The shared-blackboard run reached 1,159 GPU cycles. Adding coordination meetings reached 1,097: 62 fewer cycles, a 5.35% reduction. Lower is better.
Automated research is a method, not a single field. It applies to problems where agents can propose candidates and measure them against a defined objective. This experiment optimised a GPU kernel.
Our September 2026 research summary
Our next research direction
This September 2026 study tested scheduled meetings. Later Discovery research tested a shared direction board for comparing different approaches. Each study has its own configuration and results. Read the later Discovery study →
