The newest frontier in AI coding-agent performance may not be a bigger model but the software wrapped around it. Tech Times reports on two developments from the same week that point in this direction: Tsinghua University's HarnessPro paper and Factory's launch of Droid 1.0.

Factory's Droid 1.0

Factory released Droid 1.0, a commercial coding agent built around a technique it calls Script mode. Rather than calling tools one at a time and appending each raw result to the model's context, the model writes a small JavaScript program. That program batches tool calls, filters results in code, and passes back only relevant data. Factory's own internal benchmarks claim token usage fell by as much as 51 times in some workflows, with accuracy improving by three to seven percentage points against rival harnesses running identical models. These figures are self-reported and have not been independently verified.

What a harness is and why it matters

A harness is the runtime around an agent: which tools it may call, how it assembles context, how long it can reason, what output formats it must follow, and what verification confirms a task is done. It has often been treated as boilerplate. Earlier work called RHO (Retrospective Harness Optimization) showed that an agent could study its own execution logs, diagnose failures, and rewrite its harness. On SWE-Bench Pro, RHO lifted completion from 59% to 80% using GPT-5.5 with Codex, but per-task cost rose 42.6%, from $1.88 to $2.68. HarnessPro asks whether accuracy and cost can improve together.

The four-agent loop

HarnessPro runs several epochs, each using ten training tasks per benchmark chosen through a Determinantal Point Process to maximize diversity and difficulty. Four agents cooperate:

• The Trajectory Analyzer reviews three runs per task to separate real failure patterns from noise and to locate costly steps.
• The Evolution Planner proposes three directions, favoring cost, balance, or completion, and can carry over mechanisms that worked in other candidates.
• The Harness Editor builds each plan as a full, executable harness in an isolated Git workspace.
• The Test Synthesizer writes ten executable tests per task from the specification and an initial trajectory, then freezes them so all candidates face the same evaluation. It caught 83.3% of real failures, versus 41.7% when an LLM simply judged whether a run looked successful, because models tend to be too optimistic about their own output.

Candidates scoring below the original harness's validation pass rate are dropped, as are any beaten on both cost and accuracy by another candidate. Up to three survivors move on, labeled Lite (cost-focused), Medium (balanced), and Heavy (performance-focused), giving deployment teams a menu of trade-offs.

Results

Tests used GPT-5.5 with Codex and GLM-5.2 with Claude Code. Heavy outperformed RHO on all three benchmarks:

• SWE-Bench Pro: 80.7% versus 80.0%, at $1.97 per task versus $2.68 (26.5% cheaper), with runtime down from 520 to 412 seconds. Medium gave up about one point of accuracy for a 53.0% cost cut, to $1.26.
• Terminal-Bench 2: 80.8%, 4.5 points above RHO, with cost down from $1.09 to $0.72 (33.9%).
• ProgramBench: 69.6% versus 68.8%, with cost down from $3.59 to $2.95 (17.8%).

Averaged, Heavy gained 2.0 points and cut cost 26.1%. On a cost-per-successful-task basis, the Lite harness reached $1.11 on SWE-Bench Pro versus $3.35 for RHO, about a threefold improvement. The pattern held with the GLM-5.2 and Claude Code pairing, though two model pairings is thin evidence for broad generalization.

How savings arise

The evolved harnesses allocate reasoning effort according to task difficulty rather than simply capping tokens. On Terminal-Bench 2, Medium uses runtime hooks: a reminder at 14 tool calls to batch operations, a nudge toward final verification at 26, and a block on further tool use at 44. Reminders are attached to tool results rather than the main prompt, preserving the prefix for caching, which matters because cached input tokens cost roughly a tenth of uncached ones. The harnesses also compress tool output, fold repeated log lines, and trim developer instructions. Ablations showed these tricks do not work in isolation: adding dynamic routing or budget hooks alone to RHO dropped SWE-Bench Pro to 76.0% and 75.0%, below its 80.0% baseline. The gains depend on the whole evolution loop.

Benchmark caveats

The article also cautions against over-reading the numbers. SWE-Bench, introduced in October 2023, has faced scrutiny; a 2024 robustness analysis found roughly a third of solutions marked correct had the fix already visible in the issue text or comments, suggesting some pattern-matching rather than genuine engineering.

Implications

For the AI tooling market, the convergence of Factory's product and Tsinghua's research suggests competition is moving to the orchestration layer. Vendors with strong harness engineering, such as Factory, and those whose harnesses can be automatically tuned, may deliver better cost-performance from the same underlying models. For enterprises, cost-per-successful-task and the ability to choose Lite, Medium or Heavy profiles may become as important as raw benchmark scores.