AI Engineering

Zhu et al. (2026)

A controlled research workflow reduced reported AI research failures from 72% to 16%, counting blocked flawed studies as successful containment.

The experiment combined reproducible computation with human approval at key decisions. Its findings support preventing unreliable claims from advancing, without establishing that every run produces valid research.

72%

of unconstrained runs contained a critical failure in the main experiment.

16%

of constrained runs were classified as failures, with human rejection of flawed studies counted as successful containment.

Xiong et al. (2026)

Twelve language models managed the same subagent team, but none granted workspace permissions with better than 50% precision.

ClawArena-Team scores a text-only main agent on how well it creates, scopes, routes and schedules a fixed pool of local subagents across 41 evolving workspaces. Routing images and audio to the right specialist is nearly solved, while over-granting files and tools is the common failure.

41 scenarios

spanning 258 multi-turn rounds and 72 staged workspace updates, each scored by a shell command rather than a model judge

Under 50%

workspace-permission precision for every model, meaning subagents were granted roughly twice the files they actually read

Shilov (2026)

On 25 real fixes specified natively in Russian, the best coding agent resolves 78.7% of tasks, but the product can quietly change which model does the work.

RuBench mines recent fix commits from five live open-source frameworks, states each task in Russian like a customer request, and grades with withheld maintainer tests. Trajectory audits exposed a model-swapping fallback and GPT-5.6 fetching held-out material.

78.7%

of the 25 tasks resolved by the best configuration, Claude Code with Opus 4.8, averaged over three runs

25 tasks

mined from fix commits merged between February and June 2026, each with a Russian specification written from scratch rather than translated

Gu (2026)

Future gains in long-horizon agents depend as much on the system built around the model as on the model itself

This position paper argues that the harness, the memory, context, skill-routing, orchestration and governance layers around a foundation model, should be designed, evaluated and scaled as a first-class object, and proposes a research agenda.

Six components

make up an agent in this framework: a reasoning substrate, memory, a context constructor, a skill router, an orchestration loop and a governance layer

Context governance

is the first bottleneck: the problem is not how many tokens a model can hold but whether the right ones are assembled each turn

Fu et al. (2026)

Only EvoAgent beats the single-controller baseline across the tested task suite, while other workflows offer narrower gains.

A shared evaluation compares six multi-agent workflows using GPT-4.1. Accuracy and token use depend on the workflow and task, with extra computation offering no guarantee of better answers.

3.05 percentage points

higher pooled accuracy for EvoAgent than the single-controller baseline, using 1.27 times as many recorded tokens.

9.67 percentage points

higher accuracy for LLM-Debate on HumanEval coding tasks, despite lower accuracy across the pooled suite.

West (2026)

Free models agree with the expensive verifier when it says yes, but miss most of its real failures

Eight free candidates, five local open-weight models and three cloud free-tier routes, were retro-graded against frontier verdicts on ten real stages of agent-written code. The best caught 77% of genuine failures. Five caught almost none, while agreeing with passing work up to 98% of the time.

Vella (2026)

AI helps engineers move faster, but turns more of the job into supervision

Over six months, professional engineers reported spending less time writing code and shifting towards verification. Productivity remained positive, even as flow and other aspects of developer experience deteriorated for a growing minority.

82%

of participants reporting less time spent writing code by the second questionnaire

0.26 → 0.53

rise in the verification-minus-creation balance across the matched task-focus cohort

Lin et al. (2026)

An agent rewrites its own coding harness, and beats the engineers who used to do it by hand

Agentic Harness Engineering turns harness tuning into an automated loop. Ten iterations lift pass@1 on Terminal-Bench 2 from a bash-only 69.7% to 77.0%, past the human-built Codex harness and every self-evolving baseline.

77.0%

pass@1 on Terminal-Bench 2 after ten automated iterations, up from a 69.7% bash-only seed and above the human-designed Codex harness at 71.9%

+10.1 pp

largest cross-family gain when the frozen harness is dropped onto deepseek-v4-flash, with no re-evolution at all

Demirer, Musolff & Yang (2026)

AI coding agents triple the code developers write, but shipped software barely budges

A study of more than 100,000 GitHub developers finds that each generation of AI coding tool delivers bigger task-level gains, yet those gains shrink dramatically as they travel down the production chain toward actual releases and end users.

180%

cumulative increase in weekly commits after adopting all three generations of AI coding tools

30%

what that 180% gain shrinks to by the time it reaches actual software releases

Liu (2026)

AI coding assistants fix more code smells than they create, but introduce nearly twice the security issues they resolve

Across 304,362 AI-authored commits from 6,275 GitHub repositories, AI tools are a net positive for surface-level code quality but a net negative for bugs and security vulnerabilities, with 24.2% of all introduced issues persisting indefinitely.

484,606

distinct quality issues identified across 304,362 AI-authored commits from 6,275 GitHub repositories

89.1%

of all AI-introduced issues are code smells, the dominant but least dangerous form of technical debt

Li (2025)

AI coding agents now ship 456,000 pull requests, but their code gets rejected far more often than human work

The first large-scale dataset of autonomous coding agent activity on GitHub reveals that speed and scale are real, but acceptance rates, review dynamics, and code complexity tell a more sobering story about the gap between benchmarks and production.

456K

pull requests authored by five autonomous coding agents across 61,000 repositories and 47,000 developers

35–65%

agent PR acceptance rates, compared to 77% for human developers, a 12–42 percentage point gap

Humberd (2026)

Agency theory was built for human managers, but AI is becoming the agent nobody knows how to supervise

A new framework maps five stages of AI evolution against traditional agency mechanisms, arguing that firms need to scaffold monitoring and incentive systems now, well before AI gains full decision-making autonomy.

Routine AI

mimics rote human decisions and stays fully under human control, like a reorder system that restocks at a fixed threshold

Machine AI

adapts human-set algorithms by pulling in structured external data such as sales forecasts to improve decision inputs