Can Agent Skills Handle AI Tools for Coding? What to Expect in 2026

The 2026 Coding Reality: Productivity Gains Come With Hidden Costs

The numbers say AI coding is everywhere. The experience says something else.

Opsera's 2026 benchmark found that 90% of teams now use AI coding tools daily, with reported time-to-PR reductions of up to 58%. Yet those same AI-generated pull requests wait 4.6x longer in review and ship with 15–18% more security vulnerabilities than human-written ones.

The contradiction runs deeper than throughput. A rigorous RCT from METR gave experienced open-source developers access to AI tools and watched them slow down by 19%. Before the study, those same developers predicted AI would speed them up by 24%. After the slowdown, they still believed AI had made them 20% faster.

This perception-reality gap is the story of 2026:

• Developer favorability toward AI tools fell from over 70% to 60% in the 2025 Stack Overflow survey, with more developers now distrusting AI accuracy than trusting it.
• Senior engineers capture nearly 5x the productivity gains of juniors, widening an "execution gap" that leaves less experienced developers struggling.
• 21% of AI coding licenses go underutilized, meaning teams are paying for tools they cannot effectively deploy.
• The share of developers reporting worsened experience in at least one dimension nearly doubled, from 14% to 27%, over six months.

The frustration is consistent: solutions that are "almost right, but not quite." Engineers spend more time reviewing, correcting, and validating AI output than they save on initial generation.

"Just use AI" is no longer a strategy. The longitudinal study from Vella and Blincoe found 82% of professional engineers now spend less time writing code, with effort reallocating toward verification and supervision. The bottleneck has moved from typing to trust.

This is exactly the gap that agent skills (the SKILL.md instruction packages now flooding coding workflows) are being recruited to fill. Whether they can actually manage the verification burden, or simply add another layer of risk, is the question that will define coding in 2026. For teams looking to separate genuinely helpful skills from risky ones, a free directory like SkillzBench is being built to scan, benchmark, and verify every skill against real outcomes before it touches a production codebase.

From Writing Code to Supervising It: The Engineer's New Job

The most striking change in engineering work isn't what AI writes — it's what humans no longer do. In a longitudinal study of 95 professional engineers, 82% reported spending less time writing code, redirecting their effort toward directing, evaluating, and correcting AI output. Researchers have given this new role a name: supervisory engineering work.

That shift sounds like a promotion, but it carries a hidden cost. The same qualities that make engineers good supervisors — deep understanding of the code they're approving — are exactly what AI assistance appears to erode. In a randomized controlled trial, developers who used AI scored 17% lower on a comprehension quiz than those who hand-coded (50% vs. 67%), a gap the researchers describe as nearly two letter grades.

The detail that should worry engineering managers most: the largest score gap appeared on debugging questions — the very skill needed to validate AI-generated code. And the code being generated needs scrutiny. According to a benchmark of over 250,000 developers, AI-generated pull requests introduce 15–18% more security vulnerabilities and wait 4.6x longer in review, straining the human verification capacity teams now depend on.

Anthropic's own research team — notably, a company that sells AI coding tools — put it bluntly: productivity benefits may come at the cost of the skills necessary to validate AI-written code, especially for junior engineers whose skill development has been stunted by AI use in the first place. They urge managers to think intentionally about how they deploy these tools at scale.

The good news is that usage patterns matter. The same study found that comprehension-building interaction patterns scored 65% or higher, while heavy-reliance patterns averaged below 40%. Teams can protect verification capability deliberately:

• Require engineers to explain AI-generated code before merging it, not just approve it.
• Rotate juniors through debugging exercises and hand-coded tasks to keep foundational skills sharp.
• Budget explicitly for review capacity — AI-generated PRs already sit in review queues far longer.
• Vet the agent skills and instructions shaping AI output, treating them as a security surface rather than a convenience.

That last point is growing urgent as skills become embedded in daily workflows. SkillzBench is being built to make that vetting practical — a free directory where agent skills are security-scanned, version-pinned, and benchmarked in a sandbox so teams can see each skill's actual lift before installing it. Supervision is the job now; making sure what you're supervising is trustworthy is the part worth investing in.

Agent Skills: A New Supply Chain Hiding in Plain Sight

In December 2025, Anthropic announced Agent Skills — packaged SKILL.md instruction files for Claude Code, Codex, Cursor, and Gemini CLI — and within roughly three months, a single skill like "Find-Skills" passed 200,000 downloads according to Mitiga's analysis of the agent skill supply chain. A new layer of the coding stack has quietly assembled itself, and almost nobody is auditing it.

The reason this matters is that skills are not prompts — they are pre-baked behaviors. Mitiga's proof-of-concept demonstrated a malicious skill silently exfiltrating an entire codebase through Cursor in just four user interactions, leaving the audit log empty. Because skills bypass hardened prompt-injection defenses and reputation signals can be inflated with bot upvotes and fake downloads, stars and install counts cannot validate what a skill actually does.

OWASP now treats this as a first-class risk category. Its Agentic Skills Top 10 (AST10) is described as the first comprehensive security framework built specifically for AI agent skills, covering all major agent platforms with evidence-based mitigations. The framework exists because the underlying exposure is structural:

• Skills install with one click and inherit the agent's permissions.
• Malicious instructions can hide inside otherwise legitimate bundles.
• Popularity metrics are gameable, so reputation is not evidence.
• Audit logs may not capture skill-driven actions at all.

The scale of the surface is easy to underestimate. By the end of 2025, over 40% of code was written entirely by AI agents, and the volume of skills entering repositories keeps climbing. Yet the core question — does a given skill actually improve your agent's coding output — remains largely untested, and no marketplace currently publishes controlled, skill-on-versus-skill-off results.

That gap is exactly what verification and benchmarking are meant to close. SkillzBench is building a free, independent directory that pins every skill version to a commit SHA, runs security scans and license checks, and tests each skill in a sandbox against the same model with and without the skill — publishing the measured Lift separately from trust scores. For teams adopting AI coding tools, the practical takeaway is simple: treat every skill as a dependency, not a shortcut, and insist on evidence before installation.

Can Agent Skills Actually Improve Coding Tools? What the Evidence Says

Here's the uncomfortable truth: nobody has published a study testing whether agent skills actually make coding agents better. The honest answer comes from piecing together adjacent evidence — and it points to a conclusion most vendors won't say out loud: results depend far more on how AI is used than on which tool you install.

Start with the productivity contradictions. Opsera's benchmark of 250,000+ developers reports AI reduces time-to-PR by up to 58% (vendor research), yet METR's randomized trial found experienced developers were 19% slower with AI tools (a controlled study). Both can't be universally true — and METR itself now warns its results no longer reflect current model capabilities.

The usage pattern matters more than either headline number. Anthropic's RCT found developers who interacted with AI in comprehension-building ways scored 65%+ on follow-up quizzes, while heavy-reliance patterns averaged below 40% (the study). Skills — pre-baked instruction packages that shape agent behavior — sit squarely in this dynamic. A well-designed skill could nudge an agent toward verification and documentation; a poorly designed one could amplify the copy-paste-rely pattern that erodes debugging ability, the exact skill Anthropic found was the biggest casualty of AI assistance.

Then there's the measurement problem, which is arguably the biggest obstacle to answering the capability question at all:

• METR explicitly cautions that benchmarks can be inaccurate predictors of real-world impact, since real work involves style, testing, and implicit requirements that benchmarks don't capture.
• Popularity signals are unreliable too — Mitiga documented reputation inflation (bot upvotes, fake downloads) that makes even malicious skills look legitimate (security research).
• METR's own data shows perception gaps: developers believed AI sped them up by 20% while actually being 19% slower — self-reported "lift" cannot be trusted.

So what's a team to do? The evidence suggests the only trustworthy answer comes from controlled comparison: run the same task, with the same model, with and without the skill, and measure the difference. That's precisely the benchmarking approach SkillzBench is building — sandbox testing that publishes lift in percentage points, kept separate from ratings and popularity, because the research shows those signals don't predict real outcomes.

The realistic expectation for 2026: some skills will deliver measurable lift, some will be neutral, and some will be risky. Until independent, versioned testing becomes the norm, treat every claimed benefit the way METR treats anecdotes — as a hypothesis worth testing, not a fact worth adopting.

What to Do in 2026: A Practical Playbook for Skills and Coding Tools

The evidence from 2025 points to a clear gap: agent skills are everywhere, but no one can tell which ones actually help and which quietly introduce risk. A documented proof-of-concept showed a malicious skill exfiltrating an entire codebase through Cursor in just four user interactions, leaving the audit log empty. With Anthropic's "Find-Skills" surpassing 200K downloads within roughly three months of launch, treating skills as a security surface is no longer optional.

Treat skills as code that runs with privilege. Skills are pre-baked behaviors, not simple prompts, so a one-click install can quietly wire exfiltration into your delivery flow. Reputation signals like stars and upvotes can be faked. Pin every version to a commit SHA, scan for vulnerabilities, check licenses, and treat silencing or allowlist instructions inside a skill bundle as red flags.

Benchmark against real outcomes, not descriptions. METR's research found that developers using AI tools were 19% slower on real tasks while believing they were 20% faster, a reminder that benchmark scores and anecdotes are unreliable proxies for production impact. Test each skill in a sandbox with the same model, with and without the skill, and measure the actual lift in percentage points.

Protect debugging and verification capability on your team. An Anthropic RCT found that AI-assisted developers scored 17% lower on comprehension, with debugging questions showing the largest gap, and that comprehension-building interaction patterns scored above 65% while heavy-reliance patterns averaged below 40%. AI-generated PRs already introduce 15–18% more security vulnerabilities and wait 4.6x longer in review, so the people reviewing that code need their judgment intact.

Plan for the supervisory shift. Eighty-two percent of professional engineers now spend less time writing code, with effort reallocating toward verification and correction. Budget for review capacity, not just tool licenses, especially since 21% of AI coding licenses already go underutilized.

Stay current as skills evolve. Blocking every installation is unrealistic, so continuous re-evaluation is the path forward. Track upstream changes, re-scan, and re-benchmark as skills update. SkillzBench is building toward exactly this: a free directory that discovers skills, verifies them against commit SHAs, benchmarks lift in a sandbox, and notifies you when upstream changes alter a skill's behavior, so you can install with evidence rather than hope.