The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
The biggest change in AI pair programming was not a universal speed boost. In roughly the 2025–2026 period, coding assistants moved from suggesting the next lines in an editor toward agents that can understand a repository, modify several files, open or prepare pull requests, and work through command-line and project workflows. Developers still have to choose suitable tasks, supply context, review changes, run tests, and absorb rework.
Adoption became routine in the surveys available for this period, but productivity remained conditional. Different studies measured different populations, tasks, and definitions of “use,” so no single percentage describes how much faster every developer became.
What changed in AI pair programming?
The old “AI pair programmer” model was mainly an editor feature: type a prompt or a few characters, inspect an inline completion, and accept or reject it. That mode remains useful, especially for small, local edits. The newer workflow adds conversational context and delegated execution. An agent can inspect a codebase, propose a plan, edit multiple files, run commands, prepare a pull request, and return a change set for a developer to inspect.
Microsoft’s Q1 2026 reporting describes this expansion in GitHub Copilot as a broader platform with multiple models, task-completing agents, pull-request generation, command-line support, and integrations. The practical distinction is not that an agent is autonomous software engineering. It is that the unit of work has grown from “complete this line” to “attempt this bounded task and show me the result.”
That shift changes where judgment is needed. Developers spend less time accepting isolated suggestions and more time defining a task, deciding what context the tool may use, checking the diff, validating behavior, and deciding whether the result belongs in the product.
Adoption moved quickly, but the surveys are not interchangeable
Two JetBrains surveys show how fast usage patterns changed while also illustrating why adoption figures need dates and definitions attached. The January 2026 AI Pulse survey covered more than 10,000 professional developers. Its “regular use” measure included at least one AI tool at work; its specialized-tool measure focused on developer-oriented products.
| Survey and timing | AI use measure | Reported adoption |
|---|---|---|
| JetBrains AI Pulse, January 2026 (more than 10,000 professional developers) | Regularly used at least one AI tool at work | 90% |
| JetBrains AI Pulse, January 2026 | Used specialized AI developer tools | 74% |
| JetBrains AI Pulse, January 2026 | Reported work use of GitHub Copilot | 29% |
| JetBrains AI Pulse, January 2026 | Reported work use of Claude Code | 18% |
| JetBrains AI Pulse, January 2026 | Reported work use of Cursor | 18% |
| JetBrains Developer Ecosystem Survey, responses collected May–July 2026 (more than 15,000 professional developers) | Used AI coding agents at work at least weekly | 90% |
| JetBrains Developer Ecosystem Survey, May–July 2026 | Used AI coding agents at work daily | 68% |
The later Developer Ecosystem Survey says about 90% of respondents were in developer, programmer, or software-engineer roles. Its weekly-agent question is not the same question as the January survey’s regular-use question, so the two 90% figures should not be read as a measured increase from one identical baseline. They do show that frequent agent use was already normal in the surveyed professional population by mid-2026.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Which tools were being used?
In the May–July 2026 JetBrains results, reported work adoption was approximately 39% for Claude Code, 16% for Codex, 21% for GitHub Copilot, 12% for Cursor, and 7% for OpenCode.
| Tool | Reported work adoption, May–July 2026 | What the number means |
|---|---|---|
| Claude Code | About 39% | Share of surveyed professionals reporting work use |
| Codex | 16% | Share of surveyed professionals reporting work use |
| GitHub Copilot | 21% | Share of surveyed professionals reporting work use |
| Cursor | 12% | Share of surveyed professionals reporting work use |
| OpenCode | 7% | Share of surveyed professionals reporting work use |
These are adoption results, not independent quality scores. A product can be familiar, available through an employer, or well integrated with a team’s existing systems without producing better code on every task. The large movement in shares between the January and May–July snapshots is evidence of a fast-moving market, not a stable league table.
From suggestions to a delegated-work loop
Agentic pair programming works best as a controlled loop rather than a handoff in which a developer stops paying attention.
1. Define a bounded outcome
State the behavior, files or package boundaries, constraints, and acceptance checks. “Add pagination to the customer list, preserve the existing API, and include tests for an empty page” is a bounded request. “Improve the application” is not. Narrow scope makes an agent’s plan and resulting diff easier to evaluate.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute2. Supply repository context deliberately
Give the tool the relevant issue, coding conventions, dependency constraints, test commands, and nearby examples. Codebase understanding is now a selling point of agent workflows, but context can still be incomplete or misleading. Identify generated files, security-sensitive modules, migration rules, and components the agent must not change.
3. Let the agent plan before it edits
For a multi-file task, ask for a short plan and affected-file list first. Check whether the proposed approach matches the architecture. Correcting a wrong plan before execution is cheaper than reviewing a large, coherent-looking diff built on a false assumption.
4. Inspect the diff, not just the explanation
Read every changed file, including configuration, tests, dependency manifests, and generated output. An agent’s narrative can omit a behavior change that is obvious in the diff. Look for unnecessary refactors, duplicated logic, altered defaults, and error paths that were not part of the request.
5. Run tests and targeted checks
Use the project’s normal unit, integration, lint, type, and security checks. Add a focused test when the task changes behavior. Passing tests establish only that the checked cases pass; they do not prove that authorization, data handling, performance, or operational assumptions are safe.
Recommended Free Tools
6. Integrate in small, reversible changes
Prefer a branch or pull request that represents one coherent task. Small changes make review and rollback practical. If the agent produces unrelated edits, discard them and restate the task rather than spending review time untangling scope.
Productivity evidence does not support one universal percentage
Anthropic’s August 2025 internal analysis surveyed 132 engineers and researchers, included 53 qualitative interviews, and examined usage data. Respondents said they used Claude for about 59% of their work and perceived a 50% productivity gain. Twelve months earlier in the same internal comparison, they reported about 28% use and a 20% gain.
Those figures are self-assessments from employees of one AI company, not a controlled measurement of all developers. Anthropic explicitly noted that productivity is difficult to measure and discussed METR findings in which experienced developers working on highly familiar codebases overestimated their productivity improvement. The figures therefore describe how that internal group experienced its workflow, not a promise that another team will gain 50%.
An earlier GitHub study, published in 2022, surveyed more than 2,000 US-based developers and compared reported perceptions with anonymized usage data. It found correlations between accepting suggestions and reporting usefulness or productivity. Because it predates the current agent wave and was published by the tool maker, it is useful historical context rather than a current estimate of agent performance.
Why “faster” can be misleading
- Generation time is only one part of delivery time. Review, testing, debugging, security analysis, and rework can offset a quick first draft.
- Familiarity changes the result. An experienced maintainer may detect a subtle error quickly, while an unfamiliar contributor may spend longer validating a plausible but incorrect change.
- Baseline work differs. A repetitive documentation edit and a cross-service data migration do not offer the same opportunity for safe delegation.
- Self-reports measure perception. They can reveal whether developers feel helped, but they do not replace task-level measurements of cycle time, defects, or successful releases.
Task type matters more than a simple agent ranking
A 2026 analysis titled “Comparing AI Coding Agents” examined 7,156 pull requests involving OpenAI Codex, GitHub Copilot, Devin, Cursor, and Claude Code. In that dataset, acceptance varied substantially by task: documentation pull requests were accepted 82.1% of the time, while new-feature pull requests were accepted 66.1% of the time.
| Pull-request task in the 2026 analysis | Acceptance rate | How to interpret it |
|---|---|---|
| Documentation | 82.1% | Highest rate reported in this dataset; not a universal documentation success rate |
| New feature | 66.1% | Lower than documentation in this dataset; not a universal feature-delivery rate |
The authors’ result points to task fit: acceptance differed more by task than by most between-agent differences in the analyzed sample. It does not establish that one agent will achieve those rates in your repository, nor that an accepted pull request is defect-free. Acceptance itself can reflect local review standards, test coverage, and the cost of making corrections.
What developers actually use these systems for
Anthropic’s August 2025 internal data gives a useful picture of everyday work in its surveyed group. Fifty-five percent said they used Claude daily for debugging, 42% for code understanding, and 37% for implementing new features.
The ordering matters. Debugging and understanding often benefit from an agent’s ability to search related files, trace a call path, summarize unfamiliar code, and suggest a focused experiment. Feature implementation can also benefit, but it normally carries a larger behavioral and integration surface. Treat these percentages as descriptions of that internal sample, not a universal distribution of developer activity.
How to compare AI pair-programming tools
Popularity is a poor substitute for a workflow comparison. When evaluating two or more options, use the same task and inspect the same evidence.
| Comparison axis | Questions to ask |
|---|---|
| Suggestion mode versus delegated tasks | Does the tool mainly offer inline completions, or can it plan and execute a bounded multi-file task? |
| Context and integration | Can it work with your editor, repository permissions, issue tracker, pull-request process, and command line? |
| Task fit | How does it perform on documentation, debugging, code understanding, tests, and new features in your codebase? |
| Review burden | Are diffs small and legible? How often do reviewers request rework or find unrelated changes? |
| Evidence quality | Is a claim based on adoption, a vendor’s internal survey, self-reported productivity, or an independent task dataset? |
Run a small internal pilot with representative tasks and record completion time, review time, rework, escaped defects, and developer confidence. Keep those measures separate. A tool that drafts documentation quickly may be valuable even if it is not the best choice for a risky schema change.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Review and security remain part of pair programming
Generated code does not remove the need for tests, threat modeling, dependency review, or human approval. Check secrets and credentials, authorization paths, input validation, logging of sensitive data, dependency licenses, and migration rollback plans. For an agent that can run commands, limit permissions to the repository and environment it needs.
GitHub reported in an older developer survey that 81% of respondents expected AI coding tools to increase team collaboration. Respondents highlighted security reviews, planning, and pair programming as collaboration areas. This is an expectation captured by that survey, not evidence that collaboration outcomes have already improved everywhere.
Free tools Windows power users keep installed
One-click scans. No signup required.
Use pull requests as the control point: record the task prompt or issue, require ordinary reviewers, preserve test output, and make it clear which portions were generated or substantially assisted. Teams can then learn where delegation helps without weakening accountability.
Common failure modes and practical fixes
The agent solves the wrong problem
Symptom: The change is polished but addresses a mistaken interpretation of the issue. Fix: Ask for a plan, restate invariants, and provide a failing test or concrete example before allowing edits.
The diff is too broad to review
Symptom: Formatting, refactoring, and feature work appear together. Fix: reset the branch, split the request, and require a file or package boundary.
Tests pass but behavior is unsafe
Symptom: Existing tests are green while authorization, data exposure, or failure handling is wrong. Fix: add adversarial and boundary cases, inspect security-sensitive code manually, and test the operational rollback path.
The tool lacks crucial context
Symptom: It invents APIs, repeats an existing helper, or changes a generated file directly. Fix: point it to the canonical implementation and repository instructions, then verify imports, versions, and generation steps yourself.
Rework erases the apparent speed gain
Symptom: The first draft is fast, but review takes longer than writing from scratch. Fix: reserve agents for tasks with clear acceptance checks, measure total cycle time, and stop using a workflow that increases rework.
What this one-year change means
AI pair programming is now better described as assisted software delivery than as autocomplete. The developer’s job has shifted toward task decomposition, context management, verification, and integration while the tool can take on more of the mechanical implementation.
The change is substantial, but it is not a settled productivity law. JetBrains’ 2026 surveys show widespread and rapidly changing use; Anthropic’s 2025 internal report shows strong perceived gains in one population while warning about measurement; and the 2026 pull-request analysis shows that task category affects acceptance. Together, those findings support a practical conclusion: delegate bounded, testable work, keep review rigorous, and judge the workflow by total delivered quality rather than generated lines or adoption share.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
