PaperlayerView exact revision on arXiv

Source-linked AI summary

AgentDiscover: Autonomous Discovery with Minimal Search Scaffolding

Mahdi Farahbakhsh, Ilan Sela, Fatemeh Doudi, Vishnu Teja Kunde, Krishna Narayanan, Jean-Francois Chamberland, Dileep Kalathil

arXiv:2610.05334v1cs.AIcs.LGcs.NE

TL;DR

Fixed human-designed search algorithms choose what prior candidates LLMs see, leaving the models only to propose. AgentDiscover lets a coding agent plan and run the search with persistent graph memory and server steering. It outperforms prior discovery frameworks across evaluated domains, with reported cost and contest advantages, while its evidence has stated scope limits.

  • Problem

    In typical LLM discovery systems, a fixed human-designed algorithm selects prior candidates for the model, even though no single search strategy is best across problems and the best strategy can change during a run.

    Source: §1 · p. 2

  • Method

    AgentDiscover uses a coding agent to direct the search and propose candidates, with a graph database holding past attempts and a server evaluating submissions and steering the agent.

    Source: §1 · p. 2, §3.2 · p. 5

  • Results

    AgentDiscover outperforms prior discovery frameworks across kernel engineering, biology, algorithm design, and mathematics, and is more cost-efficient than compared frameworks.

    Source: §1 · p. 3

  • Takeaways & Limitations

    In the reported evaluations, its programs would have placed first in seven past AtCoder contests, and it matches or improves every baseline on eleven standard tasks.

    Source: §1 · p. 3

    Main-table results are single runs without variance, and the backbone model was not varied on a fixed task.

    Source: §5 · p. 9

Abstract

from arXiv · show

Frameworks that use large language models for scientific discovery typically rely on a fixed, human-designed algorithm that decides what the model sees at each step, leaving the model only the role of proposer. The model knows nothing of the search beyond what it is shown. As models grow more capable, a question arises: does a search strategy chosen by a human before the run scale better than promoting the model from proposer to planner and letting it own the search? The Bitter Lesson suggests that choosing the strategy in advance is the kind of hand-designed structure that general methods eventually outscale. We introduce AgentDiscover, in which a coding agent plans the search using its context as working memory, runs experiments, and records every attempt in a database of ideas, candidates, and their relations. This database serves as the agent's long-term memory and is structured so that the selection rules of classical algorithms such as MAP-Elites and Monte Carlo tree search each reduce to a single query, which the agent is free to use, combine, or replace. A server maintains the database and steers the agent after every submission, keeping it on course over long runs. In our experiments, AgentDiscover is more cost-efficient than existing frameworks, reaching better scores at lower cost. On tasks in kernel engineering, biology, algorithm design, and mathematics, AgentDiscover outperforms prior discovery frameworks. Its programs would have placed first among human competitors in seven past AtCoder heuristic contests, and on eleven mathematical and systems optimization tasks it matches or exceeds every baseline that uses the same model. Our code is available at https://github.com/mhdfb/AgentDiscover.

1 Introduction

Prior discovery systems let fixed, human-designed algorithms choose what an LLM sees, even as models become capable of planning and running searches. AgentDiscover gives a coding agent control of the search and a graph database for persistent memory.

1 Introduction

  • The paper situates its approach in an existing proposal-and-verification loop where programs are scored by evaluators, including systems that have produced mathematical and scientific discoveries.Examples include FunSearch’s extremal-combinatorics constructions and AlphaEvolve’s improvements on roughly a fifth of more than fifty open mathematical problems.Source: §1 · p. 2
  • Existing systems keep search selection human-designed, so the LLM only proposes candidates chosen for its context.The paper asks whether coding agents can plan the search as well as propose solutions, given that no single strategy is best across problems and the best strategy can change during a run.Source: §1 · p. 2
  • A second design question is how to retain search history so the agent can retrieve only what it needs, because reading every past attempt consumes context.The database records attempts, scores, lineage, and ideas, while retrieval costs the agent context.Source: §1 · p. 2
  • AgentDiscover lets a coding agent direct the search and propose candidates, with past attempts available through queries to a graph database.The agent can use, combine, or replace classical selection rules expressed as queries; a server evaluates submissions and returns scores with steering messages.Source: §1 · pp. 2–3
  • AgentDiscover outperforms prior frameworks across four discovery domains, would have placed first in seven AtCoder contests, matches or improves results on eleven standard tasks, and is more cost-efficient than compared frameworks.It also improves the best published bound on the first autocorrelation inequality.Source: §1 · p. 3

2 Related Work

Prior work combines LLMs with fixed evolutionary loops, coding agents, adaptive prompts and memories, or autonomous research workflows. AgentDiscover is positioned as giving an agent ownership of verifier-scored search.

LLM-based evolutionary search

  • LLM-driven evolutionary systems differ in their fixed search algorithms, while some adapt within designer-prescribed spaces; even runtime adaptation remains bounded by predefined choices.Examples include MAP-Elites and island models, novelty filtering, UCT trees, and controllers that adjust search during a run.Source: §2 · pp. 3–4

Coding agents as proposers

  • Coding-agent systems vary in how much search control they grant: some retain fixed selection rules, while CORAL gives agents control over retrieval and search direction.AgentDiscover differs from CORAL through periodic context resets, graph-based memory, agent-written tools, optional skill creation, and server steering after submissions.Source: §2 · p. 4

Adapting prompts, scaffolds, and memories

  • Other systems adapt prompts, reasoning trajectories, coding-agent scaffolds, or skill documents, while some persist semantic search state.These approaches adapt artifacts or memories rather than the candidate program alone.Source: §2 · p. 4

Autonomous research agents

  • Autonomous research agents target open-ended research judged by people or experiments, whereas AlphaEvolve-style systems use verifiers but keep search in a fixed loop.AgentDiscover combines agent-owned search with verifier scoring.Source: §2 · p. 4

3 Methodology

The methodology frames discovery as a budgeted search over programs and replaces a fixed proposer loop with a coding agent that controls its search. AgentDiscover supports this autonomy with persistent graph memory, session turnover, and server-mediated evaluation and steering.

3.1 Problem Formulation and Motivation

  • Discovery is framed as finding a high-fitness program within an evaluation budget, with a fixed evaluator scoring candidates and reporting auxiliary metrics.The existing alternative fixes the iteration in advance: an evolutionary algorithm selects candidates for the LLM, which proposes the next one.Source: §3.1 · p. 4
  • The agent queries past attempts, runs experiments, and records its work in persistent memory, making choices about candidates, directions, and strategy during the run.This shifts monitoring and control of the search to the coding agent, with AgentDiscover providing infrastructure for safe and efficient long-horizon work.Source: §3.1 · p. 4

3.2 The AgentDiscover Framework

  • AgentDiscover has a search agent, a graph database of ideas and candidates, and an optional meta agent that turns lessons from session logs into skills.Its design aims for flexible search decisions and efficiency over hours-long runs, including context and token-cost management.Source: §3.2 · p. 5
  • The search agent queries the database, plans and runs experiments, submits candidates, and can end a stalled or costly session after recording a lesson for a successor.A new session starts with fresh context and inherits the long-term memory, tools, and skills.Source: §3.2 · p. 5, Figure 3 · p. 5
  • The graph database stores ideas, candidates, sessions, lessons, and their relations; the server validates writes, scores submissions, and returns steering messages.MAP-Elites selection, which returns the best candidate per idea, is implemented as a query.Source: §3.2 · p. 5
  • Long contexts can make repeated calls costly, anchor the agent to earlier reasoning, and cause it to lose track of the task.Fresh sessions and database retrieval address context cost and anchoring; submission messages help keep the agent on course.Source: §3.2 · p. 6

4 Experiments

AgentDiscover is evaluated across discovery benchmarks, AtCoder contests, systems and mathematical optimization, and kernel building, then examined through experiments on resets, parallel agents, and meta-agent guidance.

4 Experiments

  • The evaluation spans 24 problems across systems, mathematics, AtCoder, autocorrelation inequalities, kernels, denoising, and kernel building, with comparisons at matched spend where available.The authors corrected evaluator vulnerabilities that allowed reward hacking, excluded exploited results from performance claims, and state that the corrections preserve scoring of valid solutions.Source: §4 · pp. 6–8

4.1 Results

  • AgentDiscover outperforms TTT-Discover and CORAL on every multi-domain task, and its AHC039 and AHC058 programs would have placed first in their original contests.Its AC1 upper bound of 1.50274 ranked first on the public Einstein Arena leaderboard at the time of writing.Source: Table 1 · p. 6, §4.1 · p. 8
  • AgentDiscover would have placed first in all five original AtCoder contests and outperforms the compared search frameworks on four of five tasks.With post-contest submissions included, it remains in the top five on every task and ranks first on two.Source: Table 2 · p. 7, §4.1 · p. 8
  • Across eleven systems and mathematical tasks, AgentDiscover matches or outperforms every baseline, leading on seven and tying for best on four.It uses the same backbone model as the other methods; the signal-processing comparison uses the corrected evaluator.Source: Table 3 · p. 7, §4.1 · p. 8
  • At every reported model-spend budget, AgentDiscover achieves the lowest kernel-builder cycle count; at $25, its 1,016 cycles beat every baseline at its full $118 budget.At $100, it reaches 960 cycles versus 1,098 for the strongest baseline, EvoX, a 12.6% reduction in execution cycles.Source: Table 4 · p. 7, §4.1 · p. 8

4.2 Additional Experiments

  • After stalls exceeding five candidates, context resets raise both the chance and size of an improvement, and a meta pass raises them further.Resets also limit the growing input cost per call and let the agent see the search afresh.Source: Figure 4 · p. 8, §4.2 · p. 8
  • Four parallel agents converge faster in wall-clock time than one, but show no clear difference at equal model spend.The agents share one candidate database through an MCP server.Source: §4.2 · pp. 8–9, Figure 5 · p. 9
  • 73% of guided forks, versus 33% without guidance, beat pre-fork best fitness within two candidates; guided candidates’ median fitness rose from 0.801 to 0.832.Across 15 forks, the guided run ended with higher best fitness in 10, the same in two, and lower in three.Source: Figure 5 · p. 9, §4.2 · p. 9

5 Conclusion

AgentDiscover lets a coding agent direct search while using a graph of past attempts as memory. The authors report stronger, cheaper results than comparison frameworks, with limits on run variance and backbone-model coverage.

5 Conclusion

  • AgentDiscover would have placed first among human competitors in seven past AtCoder contests and improved the best published bound on the first autocorrelation inequality.The authors also report better scores at lower cost than the frameworks they compared against.Source: §5 · p. 9
  • The authors attribute gains over fixed search to broader search visibility and multi-step experimentation, and gains over CORAL to more efficient context management.They report that context resets kept call costs low and helped agents escape stalls; a meta agent helped further.Source: §5 · p. 9
  • The graph database let fresh agents re-explore search cheaply, while agents wrote tools and ran experiments before submitting candidates.The authors say these components made AgentDiscover more cost-efficient than CORAL and other compared frameworks.Source: §5 · p. 9
  • Each main-table entry is a single run, so variance is unreported and small margins between methods should be read cautiously.The paper used Claude models and, on some tasks, GPT-6 Sol, but did not vary the backbone model on a fixed task or analyze the agents’ search strategies.Source: §5 · p. 9

AI Use Statement

Coding agents built on Claude and GPT models serve as AgentDiscover’s search and meta agents and as the backbone of every baseline. Generative AI also assisted with implementation, literature review, and readability edits, while authors reviewed the assisted work.

AI Use Statement

  • Claude- and GPT-based coding agents act as AgentDiscover’s search and meta agents and back every baseline in the study.Their use as part of the method is described in Section 3 and Appendix D.Source: AI Use Statement · p. 11
  • Generative AI tools helped implement and debug code and analysis scripts, summarize related literature, and edit the paper for readability.The authors say they did not use these tools to generate synthetic data, formulate claims or proofs, design the methodology or experiments, or interpret results.Source: AI Use Statement · p. 11
  • The authors reviewed AI-assisted work, tested assisted code, checked reported numbers against run logs and evaluator outputs, and verified citations against original sources.They take responsibility for the final content, including AI-assisted text, claims, and artifacts.Source: AI Use Statement · p. 11

A Implementation Details

The appendix adds implementation details for components described in Section 3.2 and directs readers to separate appendices for prompts, tools, and experiment settings.

A Implementation Details

A.1 Architecture and Run Lifecycle

The appendices describe AgentDiscover’s run lifecycle, graph schema, agent instructions, and execution safeguards. They also document recurring evaluator failure modes and corrections made to address them.

A.1 Architecture and Run Lifecycle

  • The orchestrator starts the database and MCP server, prepares the evaluation environment, and resumes each branch from its database iteration.Branch sessions run sequentially within a branch and in parallel across branches, then pause at a block barrier before the meta pass; branch state persists across sessions.Source: §A.1 · p. 14

A.2 Knowledge Graph Schema

  • Graph keys have uniqueness constraints, candidate keys hash program text, and candidate fitness records the evaluator’s score.The schema distinguishes logged facts from agent declarations, with some declarations checked against corresponding session-log records.Source: §A.2 · p. 14, Table 5 · p. 15, Table 6 · p. 15

A.3 Meta Pass

  • At block barriers, one meta agent reviews branch traces and outcomes, writes session lessons and guidance, and can register read-only queries for later sessions.Its container mounts transcripts read-only and excludes the evaluator, database password, solution files, and git history.Source: §A.3 · p. 15

A.4 Execution Safeguards

  • The evaluator runs candidate code separately from scoring and, by default, without network access; agents lack database credentials and use read-only graph queries.Candidate evaluation uses a separate container, though the Enroot and none backends cannot disable network access.Source: §A.4 · p. 15

B.1 Search-Agent Instructions

  • Search-agent instructions set a target, require graph exploration and an idea for each candidate, and direct agents to test, submit, learn from scores, and write a session lesson.The candidate cap is a maximum rather than a quota, and agents are told to switch ideas when progress levels off short of the target.Source: B Agent Prompts and Tools · p. 16, §B.1 · p. 16

B.2 Meta-Agent Instructions

  • The meta agent compares reasoning with outcomes and gives branches evidence-based guidance, while leaving decisions about what to try next to search agents.It may write session judgments, save read-only query tools, create skills, and update branch strategies.Source: §B.2 · p. 17

B.3 Session Briefing and Feedback

  • Session briefings and submission feedback report branch and database fitness, targets, candidate progress, time remaining, and—in later sessions—a note comparing continuation cost with starting fresh.The prompt also specifies optional internet access and a default policy forbidding online lookups; meta-written query tools are named for later sessions without exposing their query implementation.Source: §B.3 · pp. 17–18

C Evaluator Failure Modes and Corrections

  • Evaluator corrections address incomplete output validation, candidate-controlled scoring, and information leakage across the evaluation boundary.Examples include enforcing output constraints, separating candidate execution from scoring, and withholding private inputs and candidate output feedback.Source: C Evaluator Failure Modes and… · pp. 18–19

D Experimental Details

Appendix D records each evaluation setting’s model and run configuration, listing only experiment-specific variations.

D Experimental Details

  • Appendix D lists the model and run configuration for each evaluation setting, with only experiment-specific variations shown.All runs use the sandbox’s Apptainer/Singularity backend.Source: D Experimental Details · p. 19

D.1 Models, Access, and System Budget

This section specifies evaluation settings’ models, access policies, and system-level budgets, along with the comparison protocols used for the baselines.

D.1 Models, Access, and System Budget

  • Table 8 summarizes model, access policy, and system-level budget ceilings; AC1 and AC2 web access applies to both CORAL and AgentDiscover.The reasoning-effort setting is shared by the search and meta agents; Table 10 gives the meta model and schedule.Source: Table 8 · p. 19, §D.1 · p. 19
  • For the five AtCoder contests, OpenEvolve, EvoX, CORAL, and AgentDiscover use Claude Opus 5 and official system-test scores.Systems and mathematics baseline results are published results compiled by Qu et al. (2026), with all methods using Claude Opus 4.6.Source: §D.1 · p. 20
  • On multi-domain tasks, CORAL and AgentDiscover share model-spend caps, while TTT-Discover retains its published training and compute setting.The shared models are GPT-6 Sol for TriMul, MLA-Decode, and denoising, and Claude Opus 5 for AHC039, AHC058, AC1, and AC2.Source: §D.1 · p. 20
  • Baseline runs use SkyDiscover defaults for OpenEvolve, ShinkaEvolve, and EvoX, and CORAL’s default configuration, changing only model and budget settings required by the protocol.CORAL web access is also disabled in settings without web access.Source: §D.1 · p. 20
  • Candidate budgets count all submitted programs, including failed or timed-out evaluations, but only valid candidates can be reported.Model spend includes meta-agent calls when enabled; kernel comparisons use best-so-far candidates at common spend checkpoints rather than wall-clock time.Source: §D.1 · p. 20

D.2 Search Schedule and Stopping Rules

The appendix details search stopping and budgeting, meta-agent scheduling, and how model spend is estimated.

D.2 Search Schedule and Stopping Rules

  • Each setting runs one search branch except the parallel-agent ablation, which compares one branch with four; iterations continue until the shared system-level budget is exhausted.The maximum-iterations setting is a non-binding safety ceiling, and Table 9 reports maximum scheduled search work and wall-clock ceilings.Source: §D.2 · p. 20, Table 9 · p. 20
  • A session ends when the agent stalls or context carryover costs more than starting fresh; candidate and time caps limit context length, not comparison budget.Reaching the candidate cap starts another session when one remains; search and candidate evaluation have separate timeouts.Source: §D.2 · p. 20
  • Each main-table entry is the best valid candidate from one run, so main comparisons report no variance.Component ablations use five repeats per condition and report means and standard errors; paired meta-ablation runs share their first session.Source: §D.2 · p. 20

D.3 Meta-Agent Schedule

  • One meta pass runs after each completed search iteration, and Table 10 reports the meta-agent configuration by evaluation setting.The independent meta timeout can skip a pass without ending search, and no pass runs after the final iteration because no later session could use its guidance.Source: §D.3 · pp. 20–21, Table 10 · p. 20

D.4 Model-spend accounting

  • Model spend is estimated from transcript-recorded token use, pricing input, output, cache-read, and cache-write tokens at published model rates.Costs are summed across search and meta agents; estimates use list prices rather than invoice charges because runs used subscriptions.Source: §D.4 · p. 21

E Additional Results and Spend–Performance Curves

Additional results report official submissions and spend–performance curves across five AtCoder contests and three optimization tasks.

E Additional Results and Spend–Performance Curves

  • The AC1 construction was submitted to the Einstein Arena leaderboard, where it ranked first with a bound of 1.50274353 on October 3, 2026.Programs evaluated on five contests are available for AHC006, AHC012, AHC015, AHC025, and AHC041; multi-domain task programs are available for AHC039 and AHC058.Source: E Additional Results and… · p. 21

E.1 Additional Spend–Performance Curves

  • AgentDiscover finishes highest on all five AtCoder spend–performance curves, with clear leads on AHC015, AHC025, and AHC041.On AHC006 and AHC012, AgentDiscover and CORAL finish within rounding; on AHC006, CORAL scores higher on the official system test.Source: §E.1 · p. 21
  • The curves show best-so-far performance against billed model spend on the public evaluation cases for AHC006, AHC012, AHC015, AHC025, and AHC041.Figure 7 shows corresponding curves for TriMul, MLA-Decode, and denoising.Source: §E.1 · p. 21

E.2 AHC025 Scoring

AHC025 distinguishes fixed absolute scores from mutable relative standings scores. AgentDiscover has the lowest absolute score, and Figure 8 shows it reaching the contest-winner reference at lower model spend than the strongest baseline.

E.2 AHC025 Scoring

  • AHC025 relative scores and ranks come from a mutable standings snapshot, while absolute scores are fixed; Figure 8 trajectories use 50 public cases, unlike the 5,000 official cases for its reference.Source: §E.2 · p. 21, Figure 8 · p. 23
  • 19,173,103 vs 25,923,844: AgentDiscover posts the lowest AHC025 fixed absolute score, ahead of strongest baseline CORAL, with methods ordered identically under both scores.Source: Table 11 · p. 21, §E.2 · p. 21
  • At $22.34 versus $72.97 in model spend, AgentDiscover reaches the contest-winner reference sooner; at full budget, its quality is 1.96× versus CORAL’s 1.72×, with a 12.3% lower objective.Source: §E.2 · p. 21, Figure 8 · p. 23

Related papers

Loading 2610.05334v1…