AI Coding Agents in 2026: What the Research Says, How to Optimise Them, and the Best Skill Repos on GitHub
AI coding is now mainstream, but the speed gains are smaller and less certain than the hype suggests. In Stack Overflow's 2026 survey of more than 30,000 developers, 66% use coding assistants or agents. Yet the best controlled trial still cannot prove a speedup, and teams that adopt AI heavily ship bigger pull requests, wait longer for reviews and see more bugs. The developers getting real results treat the agent as an amplifier: they give it small tasks, tight context, tests to pass and a human review at the end.
This guide pulls together the latest surveys, trials and benchmarks, explains how coding agents are optimised today, lists what to do and what to avoid, and points you to the most starred GitHub repos full of ready made skills.
66%
of developers use coding assistants or agents (Stack Overflow 2026)
42%
of code is already AI generated or assisted (Sonar State of Code)
48%
of developers always check AI code before committing (Sonar)
Where the market stands in October 2026
Adoption is no longer the question. The 2025 DORA report, from nearly 5,000 technology professionals, found 90% already use AI at work. Stack Overflow's 2026 survey, published on 6 October, shows how they use it.
Adoption
How developers use AI in 2026
Stack Overflow Developer Survey 2026, 30,000+ respondents in 169 countries
Coding assistants or coding agents
General purpose chat tools
Automated agent workflows
No AI tools at all
31% of developers use AI four or more hours a day. 70% reach for an AI agent when they first tackle a technical problem, second only to web search at 83%.
Three more signals shape the picture:
- The share of AI code is climbing fast. Developers in Sonar's survey say 42% of their code is AI generated or assisted, up from about 6% in 2023.
- Trust lags far behind use. 96% of developers do not fully trust that AI code is functionally correct (Sonar), and only 6.6% would trust AI for important decisions (Stack Overflow).
- Agent preference is consolidating. For future use, 45% of Stack Overflow respondents favour Anthropic agents against 32% for OpenAI. An informal SD Times poll of 103 developers named Claude Code as the favourite tool at 41%, ahead of Cursor at 21%.
The productivity paradox: more code, not always more progress
The best evidence on speed comes from METR, which runs randomised trials on experienced open source developers. In 2025 it found developers were 19% slower with AI. Its 2026 follow up points the other way, but the confidence ranges still cross zero, so it cannot confirm a speedup.
Controlled trials
Change in task time when developers use AI (METR)
Dot is the estimate, bar is the 95% confidence range. Left of the dashed line means faster.
2025 study · 16 developers, 246 tasks
2026, returning developers · 10 developers
2026, new developers · 47 developers
METR notes that 30% to 50% of developers skipped tasks they did not want to do without AI, so the true speedup on AI friendly work is probably higher than measured. It calls the 2026 data unreliable for that reason.
Telemetry from real teams shows the other side. Faros AI tracked 22,000 developers across more than 4,000 teams and compared each organisation's low and high AI adoption periods:
Real teams
What changes when a team adopts AI heavily
Faros AI telemetry, high vs low AI adoption periods. Green is good news, orange is a cost.
Epics completed per developer
Average pull request size
Pull request review time
Bugs per developer
More work gets done, but each change is bigger, review becomes the bottleneck and more bugs slip through. DORA reached the same verdict: AI raises throughput and instability together.
DORA's summary is the most useful line in all of this research: AI is an amplifier. It speeds up teams with good habits and magnifies the problems of teams without them. DORA links the gains to seven capabilities, including strong version control, working in small batches and good internal platforms. Every optimisation below follows from that.
How good the models have become
On SWE-bench Verified, the standard test of fixing real GitHub issues, the top models now solve almost every task. That is a huge jump from about 81% for the best model a year earlier.
Benchmark
SWE-bench Verified, share of real GitHub issues solved
Steel.dev leaderboard, last updated 4 September 2026
Claude Opus 5
Claude Mythos 5
Claude Fable 5
Claude Opus 4.8
GPT-5.6 Sol
Claude Opus 4.5
The benchmark is close to saturated. Real codebases are messier than its tasks, which is why the trials above show far smaller gains than the scores suggest.
The lesson: the model is rarely the bottleneck any more. What separates good results from bad ones is the context you give the agent, the size of the task and how you check its work. Our AI research review and our study of small specialised coding models go deeper on the model side.
How coding agents are optimised today: context engineering
Prompt engineering has grown into context engineering: curating exactly what the model sees. It matters because models get worse as their input grows. Chroma tested 18 models and found every one became less reliable with longer inputs, scoring far higher on a focused 300 token prompt than on the same question buried in 113,000 tokens.
Modern agents give you six layers to control context. The rule of thumb from Thoughtworks' Birgitta Böckeler: load as little as possible up front, and add context only when a task needs it.
The context stack
Six layers, from always loaded to on demand
Skills follow the open Agent Skills standard, so the same SKILL.md works in Claude Code, Codex, Gemini CLI, Cursor, OpenCode and dozens of other agents.
The workflow that gets results
The most starred skill packs on GitHub all encode the same loop. It turns one big vague request, the thing agents handle worst, into many small checked steps, the thing they handle best.
Agent loop
Spec first, then test, build and review in small batches
This is DORA's "working in small batches" applied to agents. Each pass is small enough to review in minutes and easy to roll back.
Optimising cost: route by difficulty
Agents are input heavy. In a typical coding step the agent resends about 30,000 tokens of code and history and writes only about 1,000, so roughly 86% of the bill is input. That makes two levers matter most: keep the context lean, and do not run every step on your most expensive model.
Model routing
Cost of a 100 step agent task in rupees
Each step 30,000 input and 1,000 output tokens, Velona prices on 16 October 2026. The routed run sends 10 steps to Sonnet 5.5 and 90 to V4 Flash.
Every step on Claude Opus 5.5
Every step on Claude Sonnet 5.5
Routed: plan and review on Sonnet 5.5, the rest on DeepSeek V4 Flash
Routing cuts the bill by about 86% in this example. Check quality on your own tasks: cheap models do well on routine edits, but the planning and review steps are where a strong model earns its price.
Other cost levers that the research and the tools now support: start a fresh session for each task so history does not pile up, use prompt caching for the parts that repeat, cap max_tokens, and give each API key a budget. Our Claude Sonnet 5 pricing guide and cheapest model for coding agents have the full rupee numbers.
What to do, and what not to do
These rules come straight from the studies above and from the skill packs thousands of developers have adopted.
DoHabits that pay off
- Keep AGENTS.md or CLAUDE.md short: build commands, test commands, conventions
- Ask the agent to clarify and plan before it writes any code
- Break work into small tasks you can review in a few minutes
- Give the agent a test command and make it pass tests before it says done
- Use a fresh subagent or a second model to review the diff
- Start a new session for each new task
- Route routine steps to cheaper models and keep the strongest for plan and review
- Read the code of any skill before you install it
- Use git worktrees or branches so an agent run is easy to throw away
- Measure lead time, review time and bugs, not lines of code
AvoidHabits that cost you
- Paste huge instruction files copied from strangers into every project
- Hand the agent one giant vague task and walk away
- Merge a large AI pull request without reading it
- Let one session run for hours until the context is full
- Run every step on the most expensive model by default
- Install skills from a marketplace just because they are popular
- Give an agent production credentials or unrestricted shell access
- Load every MCP server you have into every session
- Treat a benchmark score as proof it will work on your codebase
- Skip tests because the code looks right
The most starred skill repos on GitHub
Skills are the fastest way to adopt the workflow above without writing it yourself. These are the most starred repositories focused on coding agent skills, with star counts checked on 16 October 2026. Pick one workflow pack, not three: stacking them causes the duplicate and contradictory instructions the research warns about.
GitHub stars
Coding agent skill repos by stars
- Workflow packs
- Skill libraries
- Curated lists
obra/superpowers
anthropics/skills
DietrichGebert/ponytail
addyosmani/agent-skills
ComposioHQ/awesome-claude-skills
hesreallyhim/awesome-claude-code
wshobson/agents
github/awesome-copilot
VoltAgent/awesome-agent-skills
OthmanAdi/planning-with-files
Stars show attention, not safety or quality. Read the section on skill security before installing anything.
Repo guide
What each repo gives you
A full development method as skills: brainstorming, written plans, test driven development, subagent driven development, code review and git worktrees. Works in 16 agents including Claude Code, Codex, Cursor, Gemini CLI and OpenCode.
One skill that makes the agent write as little code as possible, checking whether code is needed at all before writing the smallest version. Adds review and audit commands.
25 production engineering skills across the lifecycle: spec, planning, incremental build, TDD, context engineering, security, performance, CI and shipping.
Keeps the plan in task_plan.md, findings.md and progress.md on disk, so long tasks survive context resets.
Anthropic's official skills and the Agent Skills specification, with a template to build your own. Includes the document skills behind Claude's Word, PDF, PowerPoint and Excel output.
A plugin marketplace with 202 agents, 185 skills and 105 commands for Claude Code, Codex, Cursor, OpenCode and Copilot.
A curated list of Claude skills, resources and tools for customising workflows.
Community list of Claude Code skills, agents, status lines, tooling and plugins.
GitHub's community collection of agents, instructions and skills for Copilot.
Over 1,000 skills including official sets from Anthropic, Vercel, Stripe, Cloudflare, Supabase and others.
Skills are code: install them like dependencies
A skill can include scripts and instructions your agent will follow, so a bad one is a supply chain attack. In February 2026, Snyk scanned 3,984 skills from two public marketplaces:
13.4%
of skills had at least one critical security issue
76
skills were confirmed malicious, with backdoors or credential theft
91%
of malicious skills also used prompt injection
Snyk's warning is blunt: popularity and download counts are not security signals. Before you install any skill, read its SKILL.md and every script in the folder, avoid skills that download code at run time or ask for credentials, prefer official sources like Anthropic, Vercel or Stripe, and pin a version instead of auto updating.
A 15 minute setup for better agent results
Write a short AGENTS.mdTen to thirty lines: how to build, how to test, the conventions that matter. Nothing the agent can discover by reading the code.
Install one workflow skill packSuperpowers or Addy Osmani's agent-skills both give you clarify, plan, test and review out of the box. Read it first.
Make tests the finish lineTell the agent the exact test command and that a task is not done until it passes.
Route models by stepPoint your agent at an OpenAI compatible endpoint and use a strong model for planning and review, a cheap one for routine edits.
Review every diffUse a fresh subagent for a first pass, then read it yourself before merging.
A starting AGENTS.md can be this small:
# AGENTS.md
## Commands
- Install: pnpm install
- Test: pnpm test (must pass before a task is done)
- Lint: pnpm lint --fix
## Rules
- Plan first. List the files you will change and wait for approval.
- One small task per commit. Keep diffs under 300 lines.
- Never edit .env files or run database migrations without asking.
- Follow existing patterns in src/lib before adding new dependencies.To use cheap and strong models side by side in Cline, OpenCode or any OpenAI compatible agent, follow our guide to using Velona with coding agents, or compare the agents themselves in Cline vs Aider vs OpenCode vs OpenClaw.
Frequently asked questions
Do AI coding agents actually make developers faster?
The evidence is mixed. METR's 2025 trial found experienced developers were 19% slower with AI. Its 2026 follow up estimated 4% to 18% faster, but the result is not statistically significant. Team telemetry shows more work completed, alongside bigger pull requests, longer reviews and more bugs.
How many developers use AI coding agents in 2026?
In Stack Overflow's 2026 survey of more than 30,000 developers, 66% use coding assistants or agents and 26.2% use automated agent workflows. DORA's 2025 report found 90% of technology professionals use AI at work.
What is context engineering for coding agents?
Context engineering is curating what the model sees so it produces better results. In practice it means a short rules file such as AGENTS.md, skills that load only when needed, subagents with their own context, and hooks that run checks automatically.
What are agent skills?
A skill is a folder with a SKILL.md file and optional scripts that teaches an agent a task. The agent reads only the short description until the skill is needed. Skills follow the open Agent Skills standard and work across Claude Code, Codex, Gemini CLI, Cursor and other agents.
Which GitHub repos have the best coding agent skills?
The most starred on 16 October 2026 were obra/superpowers, anthropics/skills, DietrichGebert/ponytail and addyosmani/agent-skills. Curated lists such as ComposioHQ/awesome-claude-skills and hesreallyhim/awesome-claude-code help you find more.
Are agent skills safe to install?
Not automatically. Snyk found that 13.4% of 3,984 marketplace skills had a critical security issue and 76 were confirmed malicious. Read the SKILL.md and scripts before installing, and prefer official sources.
How can I reduce the cost of a coding agent?
Keep context lean, start fresh sessions per task, use caching, and route routine steps to a cheaper model. In our example, sending 90 of 100 steps to DeepSeek V4 Flash and 10 to Claude Sonnet 5.5 cut the cost by about 86%.
Run strong and cheap models from one key.
Claude, DeepSeek, GPT, Qwen and 300+ more, priced in rupees and paid with UPI.
Sources and method
Survey and study figures are quoted from the sources below. GitHub star counts were read from each repository page on 16 October 2026 and will keep changing. Cost examples use Velona rupee prices on the same day, with no caching discount.
- Stack Overflow Developer Survey 2026 press release
- METR 2026 productivity study summary
- DORA 2025 State of AI-assisted Software Development review
- Sonar State of Code Developer Survey 2026
- Faros AI: are AI coding assistants really saving time
- SD Times: AI coding tools in mid 2026
- SWE-bench Verified leaderboard (Steel.dev)
- Chroma: Context Rot
- Martin Fowler: Context engineering for coding agents
- Snyk: The threat landscape of agent skills