Coral Discovery · October 2026 research
Find better options
within a testing budget.
Repeating a promising approach can use up a testing budget before other ideas get a fair trial. This study tested a shared direction board that helps agents compare approaches, results and possible next steps.
Nine tasks, compared with the stronger reproduced single-agent or four-agent baseline for each task. Results are the best of three runs per condition.
What the 66% result means
The same result.
392 fewer tests.
The task was to make a small computer program perform the same calculation more efficiently. The agents proposed different versions, then ran each one through a simulator to check it and measure how much processing it needed. Each tested version counts as one test.
Discovery matched the comparison team’s best program after 202 tests instead of 594. Both teams used four agents. Discovery changed how they divided their search between different approaches.
Tests needed to match the comparison team’s best program. Fewer is better.
Each test runs one proposed version of the program. The bars start at zero and show up to 600 tests.
Why fewer tests matter
Testing alternatives takes resources. Finding a useful answer with fewer tests can leave more of your budget for other ideas. Where tests are slow or expensive, it could also mean reaching a decision sooner or at lower cost.
This study counted tests; it did not measure total time or money saved. Those depend on the cost of each test and the agents’ work between tests.
What counts as the same result?
Both programs completed the target calculation in 1,138 simulator cycles. A cycle is one step in the simulator’s processing model; fewer cycles means the program needs less processing in that model. The comparison team first reached that result at test 594; Discovery reached it at test 202.
Discovery later found a program that used 1,095 cycles, 3.8% fewer than 1,138. These are simulated processing results, not measurements on production hardware. The chart uses the selected best run of three for each team. The prior paper cited a result of 1,103 cycles.
The complete comparison
Six wins. One tie. Two losses.
The outcome compares Discovery with whichever reproduced baseline scored better on that task. The evaluation counts beneath each score are the counts reported for the selected run.
| Task / better means | Single agent | Four-agent baseline | Four agents + Discovery | Outcome |
|---|---|---|---|---|
| Kernel engineeringLower score | 1220600 evaluations | 1138600 evaluations | 1095596 evaluations | Win |
| PolyominoesHigher score | 78.99100 evaluations | 80.55100 evaluations | 91.78100 evaluations | Win |
| Signal processingHigher score | 0.7596100 evaluations | 0.7718100 evaluations | 0.7904100 evaluations | Win |
| Circle packingHigher score | 2.63597740 evaluations | 2.6358940 evaluations | 2.63598340 evaluations | Win |
| Erdos overlapLower score | 0.38126460 evaluations | 0.3821560 evaluations | 0.38098460 evaluations | Win |
| EPLBHigher score | 0.14549860 evaluations | 0.145346100 evaluations | 0.14550160 evaluations | Win |
| LLM-SQLHigher score | 0.73576990 evaluations | 0.733963100 evaluations | 0.732704100 evaluations | Loss |
| Transaction schedulingHigher score | 4608100 evaluations | 4367100 evaluations | 4566100 evaluations | Loss |
| CloudcastLower score | 61818 evaluations | 61820 evaluations | 61818 evaluations | Tie |
Circle packing and EPLB improve by very small amounts; the paper describes the EPLB margin as noise-level.
A stronger mathematical construction
The Polyominoes score rose from 80.55 in the reproduced four-agent baseline to 91.78 with Discovery. The prior record cited in the paper was 87.0.
What changed
Keep promising alternatives alive.
Sharing the best answer is useful. It can also draw every agent into refining the same approach. Discovery organises the search around distinct directions so that a different idea can keep improving before it becomes the overall leader.
- Distinguish the approaches.
A verifier reads each program change and labels the mechanism behind it. The agents record the specific change they tried and what they learned.
- Show where tests are paying off.
The board tracks improvement within each direction, gives less-tested directions an exploration allowance and keeps opening a new direction as an option.
- Guide the next attempt.
Agents finish a line of work, consult the board and claim a direction. Duplicate proposals and crowding are checked. The board advises; the agent chooses.
Conditions and scope
How the experiment was run
Each condition ran three times; the paper reports its best run, not its average. The model was Claude Opus 4.6 at medium effort. Internet access was disabled. Reading the grader was blocked except for the kernel simulator. Non-kernel results count at most the first 100 evaluations, or 20 for Cloudcast; some reported runs ended earlier. Kernel runs used approximately 600 evaluations. Crashes and timeouts count as evaluations.
Where exploration did not pay off
Discovery lost to both baselines on LLM-SQL. On transaction scheduling it beat the four-agent baseline but lost to the single agent. Spending tests on more directions can use budget that would have been better spent refining one strong approach.
Research source and comparisons
Source: Coral AI Labs, Coral Discovery: Balancing Exploration and Exploitation in Multi-Agent Autonomous Research with a Direction Board, October 2026, Table 2 and Sections 4.2–5.1. The study extends the MIT CORAL research framework. That framework is distinct from Coral AI Labs’ CoralOS product.
The battery and vessel demonstrations are separate operating simulations. These benchmark results do not establish that those simulations used this direction-board configuration.
What would a better strategy be worth?
Start with a goal, a credible starting strategy, an evaluation and limits that reflect the work.
Explore the business experiments →