We propose SciMARS, a framework for recursive task-local adaptation of multi-agent harnesses in closed-loop scientific discovery. Multi-agent systems coordinate analysis, experimental design, and verification as scientific evidence evolves. However, two problems limit their effectiveness: (i) without shared memory, agents lose observations and repeat work; (ii) fixed role procedures, handoffs, and memory policies perpetuate mistakes despite retained evidence. We address these problems with a shared evidence graph and separate component editors. SciMARS adapts roles and coordination; SciMARS + Memory adds a third editor for how agents record, retrieve, and receive evidence. Candidate probes guide composition and selection of the next harness. We evaluate scientific discovery through experimental-data replay and a digital twin. SciMARS + Memory resolves 12/18 GPT-6-Astra and 2/18 GLM-5.3 attempts on Autonomous Science Bench, compared with 8/18 and 1/18 for the base multi-agent team. On ScienceAgentBench and AppWorld, it solves 14/38 and 11/20 tasks, compared with 11/38 and 5/20 for the base team. These results support combining shared memory with harness adaptation across scientific and interactive tasks.
Figures & tables
Figure 1: SciMARS overview. (A) Execution roles use experimental feedback and optional shared memory. (B) Role and coordination editors propose candidates; SciMARS + Memory adds a memory-policy editor. Public probes guide composition and selection. The selected harness and lessons guide the next round.
Step
One round of SciMARS or SciMARS + Memory
1
Execute the current harness to obtain reference evidence.
2
Propose role and coordination edits; with memory, invoke a third memory-policy editor. Validate each scope.
3
Probe all candidates in isolated contexts with matched public inputs.
4
Compose a successor using candidate evidence and tested coordination and memory options.
5
Reuse a verified same-round execution if its complete harness matches; otherwise probe the successor.
6
Select an eligible outcome and retain its generating harness. Record fallback decisions.
Table 1: One adaptation loop, two editable surfaces. SciMARS edits P,C ; SciMARS + Memory edits P,C,M . Official evaluation follows final commitment.
Method
Resolved
MD
ML
MB
Incomplete
Total
GPT-6-Astra
Base Multi-Agent
8
6
4
0
0
18
Multi-Agent + Memory
9
4
5
0
0
18
Multi-Agent + Integrated Search
10
5
3
0
0
18
SciMARS
10
4
4
0
0
18
SciMARS + Memory
12
3
3
0
0
18
Table 2: Outcomes on Autonomous Science Bench. Each method has 18 attempts per model: three runs on each of six campaigns. MD and ML mean missing only the discovery or learning gate; MB means missing both. Incomplete attempts lack an official outcome. Repeated runs use fixed campaign fixtures. Note: To reduce execution cost, the GPT-6-Astra cohort uses a two-agent worker–reviewer team, with both agents powered by GPT-6-Astra; the two agents cover the logical execution roles described in Section 3.2 .
Figure 2: Gate outcomes on Autonomous Science Bench. Each bar summarizes 18 attempts across six campaigns. Stacked segments show resolved attempts, attempts missing only the discovery or learning gate, attempts missing both gates, and incomplete attempts without an official outcome. The right-hand labels report each method’s total resolved attempts.
Method
Total Tasks
Passes
Success
Base Multi-Agent
38
11
28.9%
Multi-Agent + Memory
38
11
28.9%
Multi-Agent + Integrated Search
38
11
28.9%
SciMARS
38
13
34.2%
SciMARS + Memory
38
14
36.8%
Table 3: Results on ScienceAgentBench. All methods are evaluated on the same 38 machine-graded tasks. Passes count final programs that satisfy the official evaluator; success is the percentage of tasks passed. Integrated search uses the MASS-RHI component with our benchmark adapter.
Method
Music
Finance
Total
Base Multi-Agent
5/10
0/10
5/20
Multi-Agent + Memory
6/10
0/10
6/20
Multi-Agent + Integrated Search
5/10
3/10
8/20
SciMARS
6/10
4/10
10/20
SciMARS + Memory
7/10
4/10
11/20
Table 4: AppWorld success by domain. Each domain contains ten tasks. Entries report successful tasks over attempted tasks.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Executed candidate
Nonzero predictions
Official MAE
Pass
Reference
0/10
78.00
No
Role-only edit
10/10
6.17
Yes
Coordination-only edit
10/10
11.80
Yes
Appendix
Table 5: Two distinct repairs of the same failed reference. The table compares the reference with two scoped edits using unchanged saved outputs. Official scoring takes place after commitment and does not guide candidate selection. The evaluator requires MAE ≤30 ; lower is better.