TL;DR: On October 4, 2026, Texas A&M University researchers published AgentDiscover, a preprint framework that promotes a coding agent from proposer to planner and reported better scores at lower cost across kernel engineering, biology, algorithm design, and mathematics.

A preprint that hands search control to the agent

According to arXiv (Cornell University), the manuscript titled AgentDiscover: Autonomous Discovery with Minimal Search Scaffolding was published on October 4, 2026 as a preprint with zero citations at publication time. According to alphaXiv, the submission was entered on 04 Oct 2026 under arXiv identifier 2610.05334 and lists Texas A&M University as the affiliation. The author list identifies Mahdi Farahbakhsh, Ilan Sela, Fatemeh Doudi, Vishnu Teja Kunde, Krishna Narayanan, Jean-Francois Chamberland, and Dileep Kalathil. According to arXiv (Cornell University), the listed topics are Evolutionary Algorithms and Applications, AI-based Problem Solving and Planning, and Artificial Intelligence Applications. The same indexing record lists funding from the National Science Foundation through RINGS: Resilient Wireless Systems for Future Uplink Traffic through Cell-Free, Loosely Coordinated Access and through FuSe2 Topic 1: Edge Reinforcement Learning with Algorithm, Architecture and Circuit Co-Design, plus support from the U.S. Army Combat Capabilities Development Command. According to arXiv (Cornell University), code for the work is available at https://github.com/mhdfb/AgentDiscover.

The question: proposer versus planner

According to arXiv (Cornell University), 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. In that arrangement, the model knows nothing of the search beyond what it is shown. The paper asks whether a search strategy chosen by a human before the run scales better than promoting the model from proposer to planner and letting it own the search. According to arXiv (Cornell University), the authors invoke the Bitter Lesson, which suggests that choosing the strategy in advance is the kind of hand-designed structure that general methods eventually outscale. According to alphaXiv, the test setting is one in which an evaluator can score a computer program, so an AI system can keep proposing solutions, testing them, and learning from the verdict. The paper then asks who should decide what to try next: a fixed search recipe designed in advance, or the coding agent itself. According to alphaXiv, AgentDiscover's key move is to let the agent choose which past attempts to inspect and which experiments to run, while a server preserves the search history and scores submissions.

How AgentDiscover organizes memory and action

According to arXiv (Cornell University), AgentDiscover uses a coding agent that plans the search using its context as working memory, runs experiments, and records every attempt in a database of ideas, candidates, and their relations. That database serves as the agent's long-term memory. According to arXiv (Cornell University), the database 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. According to alphaXiv, MAP-Elites, the strategy of retaining the best candidate for each niche, can be represented as a query that returns the best candidate for each idea. According to arXiv (Cornell University), a server maintains the database and steers the agent after every submission, keeping it on course over long runs. According to alphaXiv, a database preserves the history beyond any one agent session. According to alphaXiv, the agent decides what prior work to retrieve, what to experiment with locally, whether to write new tools, and what candidate to submit.

The four-step loop

According to alphaXiv, the loop runs in sequence as query, experiment, submit, and receive score plus steering. In the first step, the agent queries past ideas and candidates. In the second step, it experiments in its sandbox. In the third step, it submits a candidate. In the fourth step, the server evaluates the candidate and returns a score and a steering message. According to alphaXiv, the architecture includes an orchestrator, a sandboxed search agent, a shared graph-database memory, a hidden evaluator, steering, fresh sessions with shared memory, and an optional meta agent. In many earlier discovery systems, a search algorithm selects which past candidates enter the model's context, the model then proposes a candidate, and an evaluator scores it. According to alphaXiv, AgentDiscover gives those decisions to the coding agent. The figure described by alphaXiv contrasts prior work, in which a fixed algorithm selects the language model's context, with AgentDiscover, in which the search agent queries memory and experiments. According to alphaXiv, the authors state that they have not analyzed which search strategies agents use.

Kernel-builder: cycles, budget, and model

According to alphaXiv, the test case is Anthropic's kernel-builder task: optimize a parallel tree-traversal kernel on a simulated processor, minimizing execution cycles while preserving correctness. According to alphaXiv, all methods used Claude Opus 5 at low reasoning effort with the same evaluator. According to alphaXiv, at a matched $100 model-spend checkpoint, AgentDiscover reached 960 cycles, compared with 1,098 cycles for the strongest baseline, EvoX. That difference is described as a 12.6% reduction in execution cycles. According to alphaXiv, the paper reports that AgentDiscover reaches the lowest cycle count at every reported spend checkpoint. According to arXiv (Cornell University), the authors report that AgentDiscover is more cost-efficient than existing frameworks, reaching better scores at lower cost. The kernel-builder result and model-spend comparison are presented under the same budget and with Claude Opus 5 across methods. According to alphaXiv, the kernel-builder result and model-spend comparison use the same budget for all methods using Claude Opus 5.

CORAL comparison and context management

According to alphaXiv, the closest prior work is CORAL, which already lets coding agents control retrieval and search direction using shared persistent memory. According to alphaXiv, AgentDiscover's claimed distinction is principally how it manages context over long searches. CORAL stores attempts, notes, and skills as file-system artifacts. According to alphaXiv, AgentDiscover instead uses a graph database that agents query and periodically starts fresh contexts. It also allows agents to write tools and run experiments before submission, and its server returns steering after every submission. According to alphaXiv, under the same Claude Opus 5 kernel-builder protocol, AgentDiscover reached 960 cycles at $100 versus 1,148 cycles for CORAL. According to alphaXiv, the authors argue that AgentDiscover manages long-run context more cost-efficiently than CORAL. That efficiency claim is described as the authors' interpretation. According to alphaXiv, the kernel comparison changes several design features at once, so it does not isolate one cause. According to alphaXiv, CORAL already lets coding agents control retrieval and search direction, so the reported difference concerns memory form and session handling rather than agent control alone.

Broad tasks: contests, math, and systems

According to arXiv (Cornell University), on tasks in kernel engineering, biology, algorithm design, and mathematics, AgentDiscover outperforms prior discovery frameworks. According to arXiv (Cornell University), its programs would have placed first among human competitors in seven past AtCoder heuristic contests. According to alphaXiv, that contest result is referenced to page 6 of the manuscript. According to arXiv (Cornell University), on eleven mathematical and systems optimization tasks it matches or exceeds every baseline that uses the same model. According to alphaXiv, that eleven-task result is referenced to page 8 of the manuscript. According to alphaXiv, when an evaluator can score a computer program, the system can keep proposing solutions, testing them, and learning from the verdict. The broad claim covers kernel engineering, biology, algorithm design, and mathematics as the four named domains. The contest claim is framed as would have placed first among human competitors in seven past contests, not as entry in live contests. The eleven-task claim is framed as matching or exceeding every baseline that uses the same model.

Resets, stalls, and measured gains

According to alphaXiv, context resets received a more specific test when some searches stall for more than five candidates. In 21.2% of continued sessions with n equal to 650, the best score improved within three candidates. According to alphaXiv, the rate was 25.0% after a reset with n equal to 32 and 26.2% after a reset plus a meta pass with n equal to 42. The corresponding expected gains over three candidates were 0.38%, 0.86%, and 1.87%. According to alphaXiv, these are associations rather than a clean causal estimate, and the reset comparison does not by itself establish causation. The reset test distinguishes three conditions: continued sessions, reset sessions, and reset plus meta-pass sessions. The sample sizes are 650 continued sessions, 32 reset sessions, and 42 reset-plus-meta sessions. The improvement window is within three candidates after the stall or intervention. The expected-gain figures apply to that three-candidate window.

What the paper does not claim to isolate

According to alphaXiv, the authors say they have not analyzed which search strategies agents use. That limitation applies to the internal strategies selected by the agent during runs. According to alphaXiv, the interpretation that AgentDiscover manages long-run context more cost-efficiently than CORAL is the authors' interpretation. The kernel comparison changes several design features at once, so it does not isolate one cause. According to alphaXiv, the reset figures are associations rather than a clean causal estimate. The preprint status means the manuscript was published on arXiv on October 4, 2026 with zero citations at publication time. According to arXiv (Cornell University), the work is indexed under evolutionary algorithms, AI-based problem solving and planning, and artificial intelligence applications. According to arXiv (Cornell University), the listed funding includes National Science Foundation awards and the U.S. Army Combat Capabilities Development Command.

Availability and record details

According to arXiv (Cornell University), the authors' code is available at https://github.com/mhdfb/AgentDiscover. According to arXiv (Cornell University), the full author set is Mahdi Farahbakhsh of Texas A&M University, Ilan Sela of Texas A&M University, Fatemeh Doudi of Texas A&M University, Vishnu Teja Kunde of Texas A&M University, Krishna R. Narayanan of Texas A&M University, Jean-François Chamberland of Texas A&M University, and Dileep M. Kalathil of Texas A&M University. According to alphaXiv, the Texas A&M University affiliation is listed for the paper submitted on 04 Oct 2026. According to arXiv (Cornell University), the abstract states that AgentDiscover outperforms prior discovery frameworks on tasks in kernel engineering, biology, algorithm design, and mathematics. According to arXiv (Cornell University), the abstract states that AgentDiscover is more cost-efficient than existing frameworks, reaching better scores at lower cost. According to arXiv (Cornell University), the abstract states that the database records every attempt in ideas, candidates, and their relations and serves as long-term memory.

Sources referenced