AI coding agent
openJiuwen Unifies Single-Agent and Swarm Execution Semantics, Enabling Coding Harnesses to Adapt Dynamically to Diagnostics and Goals
openJiuwen brings capability composition, subagents, and multi-agent collaboration onto a shared execution foundation, then uses LSP diagnostics, context management, and explicit stopping conditions to steer long-running tasks. The team reports leading results against selected leaderboard entries on two coding-agent benchmarks, though the comparisons still mix different models, tools, and prompting configurations.
Huawei’s openJiuwen team has published a systems paper on a long-horizon coding-agent harness. Rather than adding another fixed workflow, it focuses on giving single agents, delegated subagents, and Swarms the same execution semantics. Its Rail composes tools, prompts, context, and control capabilities in sequence, using visibility gates to restrict which capabilities are available to each role. Swarm Flow provides operators including `budget`, `parallel`, `compact`, `pipeline`, stateful `agent_session`, and human intervention, allowing developers to reconfigure collaboration topologies without writing a separate execution engine for every agent architecture.
During execution, the framework keeps the base model’s policy unchanged and modifies only the state visible to the model. Context management progressively compresses history, offloads large artifacts and retrieves them on demand, and can use session affinity to coordinate the KV cache. Goal Mode classifies outcomes as continue, complete, or blocked, while separately enforcing limits on iterations, time, and resources so that agents do not misreport “budget exhausted” as task success. After code changes, type and symbol diagnostics generated by the LSP are prioritized, deduplicated, capped, and automatically returned in the next round. Cross-task Self-Reflection writes only to a retrievable experience store and does not update model weights.
The authors report 82.6% Pass@1 on the 500-task SWE-bench Verified benchmark. On Terminal-Bench 2.1, the harness scored 87.19% with GPT-5.6 Sol and 84.04% with Fable 5. The latter is closer to Claude Code’s 83.8% using the same model, suggesting that the harness may provide a benefit, but the difference is only 0.24 percentage points; differences in prompts, tools, and implementations remain uncontrolled. The paper also compares its results only with “selected” entries from the official leaderboards and does not yet provide complete cost figures, token usage, or reproducible execution images. Engineering teams should first assess whether the benchmark recipes are released in full and whether Rail’s permission gates translate into genuine process, file, and network isolation.