SwarmResearch: Orchestrating Coding Agents for Open-Ended Discovery

Autoresearch [1] demonstrated that coding agents can autonomously climb an objective to discover state-of-the-art solutions to open-ended optimization problems. However, long-running coding agents burn most of their tokens on low-level optimizations rather than exploring fundamentally different approaches. This behavior is great for making tons of mirco-optimizations, but is terrible for exploring higher-level changes.

Why do autoresearchers get stuck in a single approach? Two reasons:

• Accumulating context around one approach. Once one approach appears promising, long-running agents tend to proceed with a long-tail of low-level edits on it. Conditioned on this long history of incremental refinements, high-level shifts are unlikely especially since they requires several large program edits to beat the current best.
• Only editing one program state. Autoresearch's experimental loop exposes a single program state: all improvements are committed, regressions are reverted, and future search proceeds from the current best implementation. This loop design encourages greedy local search and makes it difficult to maintain competing high-level directions. Ideally, a harness should let search branch naturally: multiple versions of a program should evolve independently when they instantiate different ideas.

SwarmResearch solves these problems in 2 simple ways: 1) subagents with fresh context windows ideate and execute new experiments 2) all major edits have separate git branches. It's implemented as three skills compatible with Claude Code and Codex. Launch up your coding agent and give it a spin! Pretty fun to watch :)

How It Works

The harness has two primary roles: Shepherd Agents and Search Agents. The Shepherd Agent is the orchestrator. It spawns and steers waves of concurrent Search Agents and has access to global context: summaries of all Search Agent attempts, including their approach and evaluated score. Search Agents are subagents that come up with a new approach and implement, evaluate, and iterate on it. Context separation is designed to preserve diversity among Search Agents while still giving the Shepherd Agent enough information to strategically steer the population. Unlike classic orchestrator-subagent architectures, which partition complex tasks like large-scale code generation across subagents, Shehpherd Agents do not decompose the problem into precise tasks. Their role is to steer population-level search behavior by initializing and maintaining a diverse population of ideas, prioritizing Search Agents on promising ideas, and breaking out of plateaus. Shepherd Agents accomplish these goals with 3 steering mechanisms:

1. 

   Git Branch. For every new Search Agent, the Shepherd Agent creates a new git branch and worktree from either the baseline commit, a completed Search Agent’s commit, or a merge of multiple Search Agent commits; each commit includes a solution. Every Search Agent begins by reading and modifying this parent solution in its worktree, so the parent shapes the neighborhood of solutions it explores. Parent selection, through branch setup, lets the Shepherd Agent steer the population's search without prescribing specific ideas. Branches from the minimal baseline encourage new from-scratch exploration, while branches from completed agents can direct effort toward promising or underdeveloped directions. Branching expresses a wide range of search patterns e.g., concurrent Search Agent fan-out around a promising solution, or deep serial chains to persistently build on attempts.
2. Search agent type. Explorer Search Agents have fresh context windows and use the content in their worktree as context. Fresh context windows makes them less anchored to previous work and better suited for trying higher-level changes in approach. Their worktree includes a lineage-local log of attempts findings.md, which Search Agents update at every commit, to avoid repeating their ancestors' mistakes. Optimizer Search Agents inherit their parent's conversation history in addition to the worktree, giving them an accumulating history to complete long-tail refinement like autoresearch successfully does.
3. Prompts. Since Search Agents only have local context of their parent, the Shepherd Agent can share prompts to give minimal, relevant context or add directions. For example, if Explorer Agents repeat similar ideas, prompts can provide a summary of already explored ideas and instruct them to explore new ideas. Or, a prompt could share complementary findings from other Search Agents.

Comparison to Baselines

SwarmResearch's performance exceeds or matches evolutionary and multi-agent baselines on most open-ended optimization tasks, indicating more effective search of their solution spaces. SwarmResearch exceeds performance of EvoX on 14/15 tasks. It clearly exceeds CORAL on 8/15 tasks and performs similarly on 5/15 tasks. Differences between CORAL and SwarmResearch on mathematical optimization tasks Circle Packing, Erdos Min Overlap, and MMD-14-3 are small. They could be reduced by further solution polishing under higher budgets.

SwarmResearch's experiments test higher-level changes than the baselines. For the average attempt and median task, SwarmResearch changes 3.2× more lines of code than CORAL and 1.7× more than EvoX. For example, in Signal Processing, SwarmResearch's median-sized diff tries a different type of signal processing algorithm while CORAL's makes a targeted change to how signal corrections are computed. In AHC016, a median-sized diff from CORAL involves tuning thresholds, while a median-sized diff from SwarmResearch redesigns the solver into different modes applied under different conditions. On most tasks, higher-level experiments mean SwarmResearch discovers a better solution. However, within the budget, CORAL better exploited strong approaches and discovered meaningful low-level changes on the systems task EPLB and heuristics task AHC026.

Case Study: Approximate Speculative Decoding

The evolution graph at the start of the blog visualizes SwarmResearch's search trajectory on the task.

Task & Setup. Speculative decoding techniques are commonly lossless [6], [3], [7], where the target model verifies every draft token to ensure the output exactly matches what the target model would have generated. However, strict per-token verification may not be necessary for preserving downstream task accuracy, and trading some accuracy for higher speed may be desirable [8], [9], [10]. To explore this direction with SwarmResearch, our fast evaluator includes 30 total reasoning-intensive tasks from LiveCodeBench v6, AIME 2026, GPQA Diamond, and HLE-MCQ [11], [12], [13], [14]. The objective of SwarmResearch is to maximize token throughput while preserving benchmark accuracy. The baseline is a basic single-file, lossless speculative decoding implementation using rejection sampling, which is 1.68× faster than vanilla decoding. SwarmResearch uses Claude Code Opus 4.8. Before solution generation, an initial SwarmResearch run analyses the baseline from different angles without generating solutions and its findings are included in the solution generation run worktrees.

SwarmResearch's Solutions. On 100 held-out tasks, averaged across 5 seeds, SwarmResearch achieves a mean tok/s that is 4.58× faster than naive vanilla decoding, with 60.6% accuracy, while vanilla decoding's baseline accuracy is 65.8%. Vanilla autoresearch [1] achieves 1.80× speedup and CORAL [5] achieves 2.26× speedup, with accuracies of 65.8% and 58.4% respectively. These methods generated solutions for ~12 hours. For SwarmResearch, accuracy loss is due to more responses exceeding our evaluator's 16,384-token limit, despite some reasoning chains including the correct answer.

The largest throughput gains with minimal accuracy loss come from systems optimizations. The best-performing optimization is batching the target forward pass across multiple benchmark sequences. To preserve accuracy for the evaluated sampling seed, SwarmResearch uses independent RNG streams across batches for sampling independence. Autoresearch and CORAL's solutions do not use batching. The figure above shows SwarmResearch's 23rd completed Search Agent applied batching on a fresh branch from main, even after significant serial exploration on approaches like relaxed acceptance conditions. Additionally, SwarmResearch's best solution optimizes the construction of token distributions for rejection sampling by using a persistent thread pool, top-k bounded distributions, and parallel distribution construction during batching. The best solution additionally adapts the number of target-model experts. The baseline uses k=8 experts throughout, but SwarmResearch finds that k=8 is only necessary during prefill. For target-token generation, SwarmResearch uses k=4 for most tokens and increases to k=5 when the previous verification round's mean top-1 probability is less than 0.7. Unlike batching and distribution-construction optimizations, adaptive MoE expert count is not strictly lossless with respect to the original target model.

Existing inference engines [15] already implement batching for speculative decoding and optimized token-distribution processing, while adaptive expert count has been explored outside speculative decoding [16]. Other than systems optimizations, SwarmResearch explores decoding techniques that trade speed and accuracy, including relaxed draft acceptance (case study in paper discusses one variant) and skipping verification for confident tokens. Prior work explores related techniques [8], [9], [10].

Risks and Challenges. SwarmResearch's novel attempts, not in existing decoding literature as far as we know, performed poorly and we did not find them reasonable. For example, one attempt was "comonotone sampling": during rejection sampling for a draft span, it sampled a single shared acceptance threshold for all tokens in the span, rather than sampling independent thresholds per token as is typical. The model produced a run with improved performance and reasoned the change increased the acceptance rate. However, after we generated acceptance rate data and evaluated additional seeds, we found its claims were unsubstantiated. SwarmResearch generates a high throughput of ideas and experiments and careful review requires high effort. Without careful review, users may be convinced by low-quality approaches, where proposals and justifications seem attractive at first glance despite being incorrect.