Doubao-Seed-2.1-pro-0915: Multi-Agent Run Fixes 83% of 1,000 Real Issues in 36 Hours
> One hundred person-years is an entire career's worth of work.
That is the estimated effort behind Luanti, an open-source sandbox game: first commit in November 2010, 378,373 lines of code, 18,726 commits from 1,441 contributors — about 100 person-years by the COCOMO model. On September 16, Volcano Engine turned this codebase into an exam: 1,000 real historical issues covering 13 modules (rendering, networking, physics mechanics, etc.), with the official fix patches pre-isolated to simulate a repo where nobody had ever fixed them. The new Doubao-Seed-2.1-pro-0915 model dispatched multiple parallel sub-agents, ran for nearly 36 hours straight, and brought 83% of the issues to a "mergeable" engineering delivery standard.
The most telling detail hides behind that 83%: the unit changed. AI coding has long been measured by completion accuracy, unit-test pass rates, and SWE-bench-style single-problem scores. Here the yardstick is an entire issue queue plus wall-clock time — from "one problem" to "one batch", from "once" to "36 hours".
Three Upgraded Capabilities
Volcano Engine's release notes split the upgrade into three interlocking parts:
1. Multimodal coding — reading design mockups, screen recordings, 3D scenes; accuracy up 31% over the previous generation. 2. Long-horizon agent capabilities — evidence tracing, recency judgment, data verification; steadier progress on complex long tasks. 3. Token efficiency — image and video reasoning token consumption down over 30%.
The third is the easiest to overlook but economically decisive: a 36-hour run with hundreds of sub-agents burns tokens linearly with turns and tool calls. Without the cost reduction, continuous long-horizon operation doesn't pencil out. Two API entry points are available: version-locked Doubao-Seed-2.1-pro-0915 and auto-updating Doubao-Seed-Evolving, already integrated into the TRAE developer IDE and the Doubao Workspace office suite.
What Happened in Those 36 Hours
Luanti (renamed from Minetest in October 2024; mostly C++ with some Lua) was a deliberate choice: complex enough to need agents, not so large that even reading the repo is infeasible. Third-party stats show 1,254 commits and 209 contributors between March 2025 and March 2026.
The exam's isolation design is the most experimental element. Removing official patches means the model cannot reverse-engineer answers from upstream commits — it must locate root causes from code structure and issue descriptions. The 13-module spread rules out cherry-picking one easy category.
| Exam element | Design | |---|---| | Repository | Luanti, ~387k lines, mostly C++ plus some Lua | | Problems | 1,000 real historical issues | | Coverage | 13 modules incl. rendering, networking, physics | | Isolation | Official fix patches removed in advance | | Execution | Multiple parallel sub-agents | | Duration | ~36 hours | | Result | 83% reached mergeable standard |
A wording caveat: "mergeable standard" is ByteDance's own judgment — patches complete enough for review, not actually merged or maintainer-approved. The 17% that failed: no data on which modules or failure modes. That gap is exactly the slice needed to judge the orchestration's stability.
The Legacy ERP Demo
More illustrative of multimodal value: a legacy ERP system where the model received only a PC operation screen recording, a few hand-drawn business sketches, and 280,000 lines of Java source — no documentation. In under 2 hours it mapped the full procurement workflow, produced 3 mobile pages (~2,000 lines), and autonomously handled field mapping, parameter formats, and null handling, completing the order-to-warehouse-review chain.
A lot of real engineering knowledge lives in mockups, screenshots, recordings, and operators' habits rather than docs — especially in decade-old backends. If "watch recordings + read sketches" works reliably, undocumented legacy systems gain a first entry point for modernization instead of patch-based life support.
500 Sub-Agents for One Due Diligence Report
A financial-research demo dispatched 500+ sub-agents, retrieved 1,000+ web pages, cross-checked financials, capacity, fleet, and hiring data, combined satellite imagery and shipping data as external evidence, and produced a traceable due-diligence report. The emphasized skill is evidence tracing — long-horizon failures usually come from "misremembering" a nonexistent intermediate datum, not miscalculation. Unlike code fixes, though, a report's correctness may take years to verify.
What Kind of Bill Is 36 Hours
Assume a mid-level engineer needs 3 hours per medium-complexity issue: 1,000 issues ≈ 3,000 person-hours — two to three quarters of work for a small team. The difference isn't speed but parallelism, and the real balance depends on token unit price and concurrency ceilings, neither disclosed. The 30% token cut effectively pushes back the threshold beyond which long-horizon runs become economical.
Five Unanswered Questions
- Who performs third-party verification of the 83%?
- Which modules does the 17% failure concentrate in?
- Mergeable ≠ merged — how many will maintainers accept?
- Were the 1,000 issues randomly sampled or curated?
- Is 36 hours wall-clock time or aggregate compute time?
The Unit Changed; the Standard Must Follow
SWE-bench scores can be recomputed against public scripts. "1,000 real issues fixed to a mergeable standard" has public problems and environment, but the judging criteria sit with the publisher — partially verifiable, not yet a benchmark. What will really decide this path's future is whether the first teams modernizing legacy systems this way publish their failure lists, not the next model bump. A success-only scorecard and a postmortem with failure samples differ by an order of magnitude in value to those who follow.
References
1. Volcano Engine Doubao model 0915 release notes and official demos, 2026-09-16 — via Huanqiu.com coverage 2. 21st Century Business Herald, "Doubao 2.1 Pro starts writing code from images; image reasoning cost down 30%", 2026-09-16 — https://news.qq.com/rain/a/20260916A0CIXU00 3. AIBase, "Doubao 2.1 Pro 0915 Released on Huoshan Arks", 2026-09 — https://www.aibase.com/news/31085 4. BlockBeats, "Doubao 2.1 Pro's first major update in three months: 1,000 real issues, 83% mergeable", 2026-09-16 — https://www.theblockbeats.info/flash/367392 5. DAMO developer matrix AI news daily, 2026-09-17 — https://damodev.csdn.net/6aab20dacf836948fcd2879b.html