Paper of the Week — SWE-Router: Routing in Multi-Turn Agentic Software Engineering Tasks
SWE-Router cuts LLM agent costs by routing easy GitHub issues to cheaper models mid-trajectory, not just at task start, using execution signals as routing features.
SWE-Router: Routing in Multi-Turn Agentic Software Engineering Tasks
Seongho Son, Sangwoong Yoon, Jiahua Tang, Shuhan Wang, Lorenz Wolf, Ilija Bogunovic. Published 2026-07-01. arXiv:2607.00053
One sentence summary
SWE-Router reduces agentic software engineering costs by 40–60% with minimal accuracy loss by routing GitHub issues to cheaper models based on signals collected mid-trajectory, not just from the initial task description.
Why this paper
Most LLM routing research assumes you can pick the right model before you start. In multi-turn agentic tasks like software engineering, you often don’t know how hard a task really is until you’re several tool calls deep — this paper tackles that gap directly.
What they did
The authors built a router for SWE-bench-style agentic coding tasks that observes signals accumulated during early trajectory steps — things like which files were touched, how many tool calls were made, and whether tests passed — before deciding whether to continue with a cheap model or escalate to a frontier one. Unlike static routers that classify from the issue description alone, SWE-Router can switch models mid-task once it has execution evidence. They evaluated across several open and proprietary models on SWE-bench Verified.
Key findings
- Routing on trajectory signals outperforms routing on task descriptions alone by 8–12 percentage points in accuracy-cost tradeoff
- Cost reductions of 40–60% relative to always using a frontier model, with less than 3% drop in resolve rate
- Early trajectory signals — specifically test outcomes and file-change patterns after the first 3–5 steps — are the strongest routing predictors
- A simple gradient-boosted classifier over these features beats learned neural routing approaches in this setting
- Routing thresholds are tunable: tighter budgets favor more aggressive early escalation
Why it matters for practitioners
If you’re running coding agents in production — bug triage, PR review, automated fixes — you’re almost certainly over-spending on frontier model calls for issues that Devstral or a smaller Qwen3 variant could handle. SWE-Router gives you a principled framework for making that call dynamically, after you’ve seen some execution evidence, rather than guessing upfront from issue text alone.
What you can use today
- The core insight is immediately actionable: instrument your agentic harness to log tool-call counts, file paths touched, and test pass/fail after the first few steps, then train a lightweight classifier to predict whether escalation is needed
- The paper’s feature set (trajectory length, test signals, file-change entropy) maps cleanly onto any SWE-bench-style harness like SWE-agent or OpenHands — you don’t need to redesign your stack
- If you’re already logging agent traces, you likely have enough data to prototype a routing classifier in a few days; start with a simple threshold on test-pass rate at step 5 as a zero-shot baseline before adding the full feature set