SwarmResearch¶

external · status: active · focus: end-to-end · discipline: computer-science · started: 2026

Project page: https://github.com/SwarmResearch/SwarmResearch

Licence: none

No licence declared

The repository declares no licence, so its code cannot be reused, modified or redistributed without the maintainer's permission. RISE describes and links to the project; nothing from it is reproduced here.

Source: projects/landscape/swarmresearch.yml

Positioning¶

An orchestrator-subagent harness for open-ended discovery, from a four-author group at UIUC. A Shepherd Agent holds the global picture and spawns waves of Search Agents by launching non-interactive Codex or Claude Code sessions: Explorers start from a fresh context and attack a new approach, Optimizers fork a parent's conversation history and refine an existing solution, and the Shepherd steers the population through three levers — parent selection, agent type and prompt. Every major edit lands on its own git branch. Sits beside AlphaEvolve and CORAL in the discovery-not-drafting corner of the landscape: it produces algorithms and solutions, not manuscripts.

Distinctive contribution¶

It ships as agent skills, not as a framework — three Markdown skills (swarmresearch, swarmresearch-explorer, swarmresearch-optimizer) in .claude/skills and .codex/skills, with the README noting that other coding agents are supported by changing the spawn commands, so the prompts are the implementation. The architectural claim is about breadth over depth: a single long-running agent converges on one high-level approach and then grinds on low-level edits, whereas branch-isolated waves keep competing approaches alive and scale parallelism adaptively with search depth. Reported to match or beat SOTA LLM-guided evolution and multi-agent baselines on 13 of 15 tasks at a $50-per-task budget.

Evaluation scores¶

Dimension	Score (0–3)	Note
Lifecycle coverage	1	Three adjacent stages — propose an approach, implement it, score it against the evaluator. No literature access, no write-up, no review of scholarship.
Autonomy level	3	The Shepherd spawns non-interactive Codex/Claude Code sessions and steers them without human gates; the paper's reference run put 50 agents through ~3 hours unattended on a $50 budget.
Architectural transparency	2	The three skills are plain Markdown in .claude/skills and .codex/skills, so the orchestration prompts and spawn logic are fully readable — but the repo holds only those plus assets/ and one spec_dec_example, has no license, and the paper's experiment harness lives in a separate reproduction repo.
Inputs supported	0	One narrow input form — a working directory holding prompt.md, an executable evaluator and an optional baseline — with no literature retrieval and no dataset connectors of any kind.
Outputs / reproducibility	2	Competing approaches persist as first-class git branches, so the search history is inspectable rather than collapsed; but there is no run manifest or seed control, and the reported 13/15 result is not reproducible from this repository.
Internal evaluation	2	The paper evaluates 15 tasks across three families — 5 math (circle packing, Erdős min overlap, signal processing, MMD-14-3, 3rd autocorrelation), 5 ADRS systems tasks (EPLB, LLM-SQL, txn scheduling, Cloudcast, PRISM) and 5 ALE-Bench-Lite AtCoder heuristics — against CORAL, EvoX and fixed-scaling baselines; arXiv preprint, not peer-reviewed or third-party replicated.
Openness	1	No LICENSE file and the GitHub API reports license: null (verified 2026-09-08), so the skills are readable but not legally reusable; a run also needs a paid Codex or Claude Code subscription, budgeted at $50 per task in the paper.
Maturity / traction	1	16 stars, 3 forks, 13 commits, nothing pushed since 2026-07-07; a single-team research prototype attached to one preprint, with no release or packaging.
Cross-family policy	1	Skills ship for two runtimes and the paper's own configurations mix families across roles (Claude Code for the swarm, Opus 4.6 for EvoX, Minimax-M2.5 subagents under a Sonnet-4.6 orchestrator), so cross-family setups are supported — but nothing requires one, and there is no LLM reviewer role to pair, because scoring is delegated to a deterministic external evaluator.
Runtime assurance	1	One in-flight check: every candidate is scored by the user-supplied external evaluator, which the README frames as reward-hacking prevention, and branch isolation stops a failed approach contaminating others. No second integrity gate, and no manuscript to audit.
Cross-platform portability	2	Skills-as-Markdown for two agent runtimes (Claude Code and Codex) with a documented path to others by changing spawn commands, and the paper exercises several model families; only two runtimes actually ship, so short of the framework-agnostic band.

Scored on 2026-09-08. See the evaluation rubric.

Tags¶

Pipeline stages: hypothesis-generation code-generation data-analysis

Architectural features: multi-agent tool-use iterative-loop artifact-versioning

Inputs: problem-description external-evaluator baseline-solution

Outputs: solution-implementations git-branch-history evaluator-scores

Knowledge sources: user-supplied-evaluator

Limitations¶

No LICENSE file (verified 2026-09-08 via the GitHub API and a root directory listing) — the skills are readable but reuse terms are undeclared.
Requires the user to supply an executable evaluator, which confines it to problems with a cheap machine-checkable objective. Most empirical economics questions have no such scorer: the quantity of interest is an identified effect, not a number a script can rank.
Scope is algorithm and heuristic design, not scholarship — no literature access, no citation handling, no write-up stage, so nothing here produces or checks a paper.
The reported 13/15 result cannot be reproduced from this repository: it holds the skills and one speculative-decoding example, while the paper's experiment code sits in a separate reproduction repo.
Expensive and slow at the demonstrated scale: $50 per task in the benchmark, and the speculative-decoding case study ran roughly 11 hours per phase on Claude Code Opus 4.8.
16 stars, 3 forks, 13 commits, unchanged since 2026-07-07 — the adaptive-parallelism claim rests entirely on the authors' own runs.

Papers describing this project¶

SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery — Virk, Y., Edds, Z., Xia, C. S., Zhang, L. (2026). arXiv. arXiv:2607.02807

virk2026swarmresearch (BibTeX)