TL;DR: Texas A&M researchers introduced AgentDiscover on Oct. 4, 2026, a discovery framework that lets a coding agent plan its own search instead of following a fixed human-designed search recipe.

A planner, not just a proposer

Most large-language-model discovery frameworks assign the model a narrow job. A human-designed algorithm decides what information appears in the model's context at each step, and the model responds by proposing a candidate. The model does not see the broader search beyond the items selected for it, according to the alphaXiv summary of the paper.

AgentDiscover, from Mahdi Farahbakhsh, Ilan Sela, Fatemeh Doudi, Vishnu Teja Kunde, Krishna Narayanan, Jean-Francois Chamberland and Dileep Kalathil at Texas A&M University, changes that division of labor. The coding agent uses its context as working memory, conducts experiments, and records every attempt in a database containing ideas, candidates and relations among them, according to the Robotics Center Research index of the work.

The central research question is whether a search strategy selected by people before execution improves with scale more effectively than elevating the model from proposer to planner and giving it responsibility for the search. The authors invoke the Bitter Lesson, the observation that advance selection of strategy is the type of hand-crafted structure that general methods later exceed, as motivation for testing agent-led planning.

The paper was submitted on Oct. 4, 2026, and listed as arXiv identifier 2610.05334. Its code was released at https://github.com/mhdfb/AgentDiscover.

How the database replaces fixed search code

The long-term memory in AgentDiscover is a database rather than a short prompt history. Every attempt is stored with its idea, its candidate implementation and links between entries. The agent consults that store across sessions, so work persists after an individual session ends.

The database is organized so that selection rules from established search methods become simple lookups. The selection rules of classical procedures such as MAP-Elites and Monte Carlo tree search each reduce to a single query, according to the alphaXiv abstract. The agent is free to invoke those queries, combine them, or replace them with a different approach.

MAP-Elites is described in the paper as retaining the best candidate for each niche. In AgentDiscover that behavior is represented as a query that returns the best candidate for each idea, according to the alphaXiv overview. The agent may therefore reproduce niche-preserving search when useful, without being locked into it.

A server maintains the database and interacts with the agent throughout extended runs. After every submission, the server steers the agent, a mechanism intended to keep the search on course over long sequences of attempts, according to the same overview.

The four-step loop

AgentDiscover operates through a repeated cycle of query, experiment, submit and feedback. First, the agent queries prior ideas and candidates from the shared store. Second, it experiments inside its sandbox. Third, it submits a candidate. Fourth, the server evaluates the candidate and returns a score plus a steering message, according to the alphaXiv page.

That sequence differs from earlier discovery systems in which a fixed algorithm selects the context shown to the language model. In the earlier pattern, the model proposes a candidate and an evaluator scores it. In AgentDiscover, the search agent queries memory and performs experiments before proposing, as illustrated in the paper's workflow comparison.

The agent's local powers are broader than proposal alone. It decides which prior work to retrieve, what to test locally, whether to create new tools, and which candidate to submit, while the database preserves history beyond any one session, according to the alphaXiv summary.

The architecture includes an orchestrator, a sandboxed search agent, shared graph-database memory, a hidden evaluator, steering messages, fresh sessions that share memory, and an optional meta agent. The numbered cycle links query, experiment, submit, and receipt of score plus steering.

The authors state that they have not analyzed which search strategies the agents actually use in practice. That leaves open whether agents favor familiar queries, mix them, or develop different selection patterns during runs.

Kernel-builder test at matched spend

When an evaluator can score a computer program, an AI system can repeatedly propose solutions, test them and learn from the result. The paper's detailed test uses Anthropic's kernel-builder task, which asks for optimization of a parallel tree-traversal kernel on a simulated processor, minimizing execution cycles while preserving correctness.

All compared methods used Claude Opus 5 at low reasoning effort with the same evaluator, according to the alphaXiv overview. The comparison therefore holds model and scoring constant while varying the discovery framework and its search management.

At a matched $100 model-spend checkpoint, AgentDiscover reached 960 cycles. The strongest baseline, EvoX, reached 1,098 cycles at the same spend, according to the same overview. That difference equals a 12.6% reduction in execution cycles.

The paper reports that AgentDiscover reaches the lowest cycle count at every reported spend checkpoint in that kernel-builder protocol. The result is presented as evidence that the framework reaches better scores at lower cost, a claim stated in the abstract indexed by Robotics Center Research.

Cost efficiency is a core assertion of the work. The abstract states that AgentDiscover is more cost-efficient than existing frameworks, reaching better scores at lower cost across its experiments. The kernel-builder numbers provide the most specific spend-matched illustration of that claim.

CORAL, EvoX and context management

The closest prior work identified in the paper is CORAL, which already gives coding agents control over retrieval and search direction through shared persistent memory. The claimed distinction for AgentDiscover concerns principally how context is managed over long searches, according to the alphaXiv overview.

CORAL stores attempts, notes and skills as file-system artifacts. 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 that comparison.

Under the same Claude Opus 5 kernel-builder protocol, AgentDiscover reached 960 cycles at $100 of model spend, compared with 1,148 cycles for CORAL, according to the alphaXiv summary. The gap to CORAL is larger than the gap to EvoX in the reported checkpoint.

The authors interpret the kernel comparison as showing that AgentDiscover manages long-run context more cost-efficiently than CORAL. That interpretation is theirs, and the comparison changes several design features at once, so it does not isolate a single cause for the difference.

The broader experimental claim extends beyond kernels. On tasks in kernel engineering, biology, algorithm design and mathematics, AgentDiscover outperforms prior discovery frameworks, according to the abstract carried by Robotics Center Research.

Contests, math tasks and stalled searches

Two additional results describe breadth. The paper reports that its programs would have placed first among human competitors in seven past AtCoder heuristic contests, according to the abstract. On eleven mathematical and systems optimization tasks, it matches or exceeds every baseline that uses the same model, according to the same abstract.

The ArXiv TLDR summary of the paper lists four key contributions: promoting language models from proposers to planners so they own the search strategy; using a database as long-term memory structured for flexible query-based algorithm selection; greater cost efficiency with better scores; and outperformance of prior frameworks across kernel engineering, biology, algorithm design and mathematics.

Context resets received a separate measurement. Some searches stall for more than five candidates without improvement. In continued sessions without a reset, the best score improved within the next three candidates in 21.2% of cases, based on 650 sessions, according to the alphaXiv overview.

After a reset with a fresh context, the improvement rate within three candidates was 25.0%, based on 32 sessions. After a reset plus a meta pass, the rate was 26.2%, based on 42 sessions, according to the same overview.

Expected gains over three candidates followed the same ordering. The corresponding expected gains were 0.38% for continued sessions, 0.86% after a reset, and 1.87% after a reset plus a meta pass. Those figures are associations rather than a clean causal estimate, and the reset samples are much smaller than the continued-session sample.

The workflow figure in the paper contrasts prior work, where a fixed algorithm selects the language model's context, with AgentDiscover, where the search agent queries memory and experiments. The numbered cycle of query, experiment, submit and score-plus-steering is shown alongside the orchestrator, sandboxed agent, shared graph memory, hidden evaluator, steering, fresh sessions and optional meta agent.

Together, the reported pieces describe a minimal-scaffolding approach: preserve every idea and candidate in a queryable graph, let the agent choose retrieval and experimentation steps, evaluate each submission with a hidden scorer, return steering, and restart context when searches stall. The code release lets others inspect the server steering mechanism, database queries and sandbox tools that implement that approach.

Sources referenced