icat-agent: Multi-agent scaffolding boosts issue resolution by 8.3% on SWE-bench Pro
New decentralized agent framework unlocks 67.4% fix rate with GPT-5.4-xhigh, costing $1.18 less per instance.
icat-agent is a decentralized, multi-agent scaffolding designed to resolve complex software issues with ambiguous, incomplete descriptions. Unlike prior systems that rely on a shared context window, icat-agent uses synchronous, event-based message passing to avoid context degradation and poisoning. It also incorporates a rubric-based quality check to dynamically pivot its workflow: for high-quality issues, it initiates parallel patching and validation; for low-quality ones, it first performs preliminary exploration before fixing. This adaptive approach significantly boosts success rates on challenging software engineering benchmarks.
Evaluated on SWE-bench Verified and SWE-bench Pro, icat-agent consistently outperforms prominent baselines including SWE-agent, mini-SWE-agent, and Claude Code—all using the same underlying models. Gains range from 3.6–8.4% on SWE-bench Verified to 6.3–18.5% on SWE-bench Pro. Moreover, icat-agent is computationally efficient, reducing average cost by $1.18 per instance compared to multi-agent Claude Code. Most strikingly, the icat-agent + GPT-5.4-xhigh combination achieves 67.4% on SWE-bench Pro, surpassing the previous state-of-the-art (mini-SWE-agent + GPT-5.4-xhigh at 59.10%) by 8.3 percentage points, demonstrating that superior scaffolding can unlock latent model capabilities.
- icat-agent uses event-based message passing instead of shared context, preventing context poisoning in long-horizon workflows.
- Adaptive workflow: parallel patching for clear issues, preliminary exploration for ambiguous ones; improves fix rates by 3.6–18.5% over baselines.
- With GPT-5.4-xhigh, icat-agent achieves 67.4% on SWE-bench Pro, outperforming previous best by 8.3 percentage points, while costing $1.18 less per instance than Claude Code.
Why It Matters
Better scaffolding, not bigger models, drives major gains in automated bug fixing, slashing costs and boosting AI reliability for enterprise codebases.