跪拜 Guibai
← All articles
Agent · Programmer · Artificial Intelligence

AI Coding Agents Are Fast. The Harness That Controls Them Is What's Missing.

By 纯爱掌门人 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Teams that treat AI coding tools as a drop-in productivity booster are already seeing review backlogs, larger failure-prone PRs, and net-zero delivery gains. The organizations getting real value are the ones investing in the control systems around the agent — and that gap will widen as models become commodities.

Summary

Three years of industry data from DORA, Stack Overflow, Faros AI, and METR converge on a single finding: AI makes individual developers feel faster, but team-level delivery throughput and stability often stall or regress. PRs get larger, review times balloon, and bugs per developer tick upward. The bottleneck has shifted from writing code to clarifying requirements, verifying correctness, and integrating changes safely.

What's missing is a Harness — the organizational layer of task specs, project knowledge, tool permissions, sandboxes, test gates, and failure-driven improvement loops that surrounds a raw agent. OpenAI, Spotify, and LangChain have all demonstrated that tuning this external system, not just swapping models, is what unlocks reliable AI-assisted delivery.

A 90-day blueprint for building a minimum viable Harness starts with baselining real task metrics, turning tribal knowledge into executable maps, risk-tiering autonomy, embedding verification into the agent's own loop, and isolating permissions. The end goal is not more code, but auditable evidence that AI-generated changes are correct and safe.

Takeaways
Google's 2025 DORA report found AI improves delivery throughput but still harms delivery stability, acting as an amplifier of existing team maturity.
Faros AI observed that high AI-usage teams merged 98% more PRs but saw review time increase 91% and bugs per developer rise 9%, with no significant overall performance gain.
METR's 2025 study of experienced open-source developers found AI made them 19% slower on average on familiar large projects, despite their own perception of being faster.
Anthropic's analysis of 400,000 Claude Code sessions shows humans still make 70% of planning decisions; domain expertise directly determines how much work an agent completes per instruction.
61.38% of agent-generated PRs in a 2026 GitHub study received no human review at all, and 58.77% of reviewed PRs were only checked by other automated agents.
OpenAI built a dedicated Harness with structured docs, isolated workspaces, and automatic verification to achieve its 1-million-line Codex project, and warns results are not replicable without equivalent investment.
LangChain improved an agent's benchmark score from 52.8% to 66.5% without changing the model, only by adjusting prompts, tools, and the verification loop.
Spotify's 76% PR frequency growth with AI rests on years of prior investment in Backstage, a unified tech stack, automated maintenance PRs, and isolated agent runtimes.
The core metric is whether AI-saved human time exceeds the verification, rework, and incident costs it creates.
Conclusions

AI coding productivity data splits cleanly along one axis: studies measuring isolated, greenfield tasks show large gains; studies measuring end-to-end delivery on real codebases show flat or negative results. The field has been measuring the wrong thing.

The finding that 61% of agent PRs receive zero human review, and most of the rest are reviewed only by other bots, suggests the industry is sleepwalking into an accountability vacuum where no human has actually read most merged changes.

OpenAI's decision to stop using SWE-bench Verified due to contamination and later retract SWE-bench Pro recommendations undermines the entire leaderboard-driven model-selection culture that many teams rely on.

Spotify's results are often cited as an AI success story, but its 250 million automated maintenance PRs predate LLM agents — the platform engineering was the real force multiplier, and the agent is just the latest consumer of that infrastructure.

The shift from 'model competition' to 'system competition' implies that proprietary internal task sets, failure logs, and permission architectures will become harder-to-replicate moats than model access itself.

Concepts & terms
Harness
The full working system surrounding an AI coding agent: task specifications, project knowledge, tool permissions, runtime sandboxes, test gates, approval rules, process logs, and failure-driven improvement mechanisms. Distinct from the agent itself, which is just the code-generating component.
SWE-bench Verified / Pro
Public benchmarks used to evaluate AI coding agents on real-world GitHub issues. OpenAI withdrew support for both in 2026 after discovering test flaws, training data contamination, and task errors affecting roughly 30% of the Pro set.
DORA AI Capabilities Model
Google's framework identifying key capabilities for effective AI-assisted delivery: small batch sizes, strong version control, internal platforms, and accessible internal knowledge, among others.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗