The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →AI coding agents can now take on multi-step work, not just suggest lines of code. That shifts some developer effort from carrying out each step to setting goals, supplying context, checking results and deciding what is safe to accept. But adoption and reported productivity gains do not prove that developer jobs are either safe or doomed: the evidence available here does not establish an economy-wide employment effect.
What does “agentic AI” mean in software development?
A coding assistant typically responds to a prompt with a suggestion, such as a code completion or an answer. An agent can pursue a goal through multiple actions: inspect a codebase, edit files, run tests, respond to failures and continue working. The distinction is the amount of execution the system undertakes, not whether a human has disappeared from the process.
As an Amazon Associate I earn from qualifying purchases.
In Anthropic’s analysis of around 400,000 interactive Claude Code sessions involving around 235,000 people from October 2025 through April 2026, observed tasks included building and fixing software, testing, coordinating agents, operating software, understanding systems, planning changes, analyzing data and writing documentation. Anthropic summarized the observed division of labor this way: “People decide what to build, and the agent decides how to build it.” That describes use of one vendor’s product in its analyzed sessions; it is not a universal rule for every agent or development team. Anthropic’s analysis of Claude Code use
Is AI taking software developers’ jobs?
The available studies do not answer that economy-wide question. They examine tool adoption, work practices, reported experience or particular experiments—not a causal forecast of how many developer jobs will exist. It would be inaccurate to turn high adoption into proof of job losses, or reported benefits into proof that employment is unaffected.
#1 Best Overall
There is evidence that developers themselves worry about displacement. Anthropic’s December 2025 internal study surveyed 132 of its engineers and researchers and conducted 53 in-depth interviews. Participants reported productivity gains and broader task coverage, while also raising concerns about displacement, keeping technical skills sharp, supervising outputs and collaboration. Anthropic notes that its employees had early access to the tools and worked at an AI company, so their experience cannot stand in for the profession as a whole. Anthropic’s study of work at the company
How widely are developers using coding agents?
In a weighted survey of more than 15,000 professional developers worldwide, JetBrains reported that 90% of respondents used AI coding agents at work weekly and 68% did so daily. The survey was conducted from May through July 2026. These are estimates from survey respondents, not a census of developers or a measure of how much work agents completed. JetBrains’ 2026 coding-agent adoption survey
GitHub’s 2025 survey update offers a different measure: more than 97% of 2,000 respondents said they had used AI coding tools at work at some point. That question did not ask how often they used the tools, and GitHub cautions that reported use does not mean an employer approved it. The figure therefore cannot be directly compared with JetBrains’ weekly and daily usage estimates. GitHub’s survey update
What does the evidence say about productivity?
There is no single productivity number that applies to every developer, task or team. The evidence includes self-reported perceptions, observational product data and field experiments, which answer different questions.
Developers’ reported experience
Microsoft Research’s SPACE study collected survey responses from more than 500 developers. It uses five dimensions—Satisfaction, Performance, Activity, Collaboration and Efficiency—and reports that developers broadly perceived AI as useful, particularly for routine work. Its summary says effects varied with task complexity, individual usage patterns and team adoption; it found less evidence of a collaboration effect and points to organizational support and peer learning as relevant factors. These are reported experiences, not a fixed productivity lift for all developers. Microsoft Research’s SPACE study
GitHub’s survey respondents reported benefits involving code quality, efficiency, test generation, onboarding and understanding codebases. Those are respondents’ perceptions, not proof that the tools caused measurable improvements in every setting. GitHub’s survey update
What field experiments can—and cannot—show
Microsoft Research describes randomized field experiments at Microsoft, Accenture and an anonymous Fortune 100 company. In each, a randomly selected subset of developers received an assistant that suggested code completions. The page establishes the experimental design; it does not support a universal productivity effect for all AI agents or software work. Code-completion experiments also do not, by themselves, establish the impact of multi-step agents that can operate tools and pursue broader tasks. Microsoft Research’s field-experiment study
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Which parts of a developer’s work are changing?
Agentic systems can shift effort across the development process rather than simply shorten the act of typing code. Depending on the system and workflow, a developer may spend more time describing intended behavior, choosing context, reviewing a proposed plan, monitoring execution and checking that the finished change meets its requirements.
- More execution can be delegated: agents may carry out bounded sequences such as exploring files, making edits and running tests.
- More work may happen outside code authoring: observed Claude Code sessions included testing, analysis, software operation and documentation as well as building and fixing.
- Human decisions still shape the result: someone must determine what should be built, what constraints matter and whether an output is acceptable.
The exact balance varies. A small, well-specified change with reliable tests is different from unfamiliar code, an ambiguous product request or a high-risk deployment. The evidence does not show that every developer’s day is changing at the same pace or in the same way.
How much autonomy should an agent have?
Autonomy is a work-design decision: should a system suggest an action, perform a bounded task for approval, or continue through multiple steps with fewer interruptions? The answer can depend on the task’s impact, reversibility, available tests and how well a reviewer can inspect the result.
Microsoft Research studied acceptable autonomy boundaries across software-engineering work with 448 professional developers at Microsoft. The study establishes that developers’ boundaries are an empirical question; it does not show that all developers accept the same level of automation. Microsoft Research’s study of autonomy boundaries
Microsoft WorkLab’s 2026 Work Trend Index combines anonymized Microsoft 365 signals with a survey of 20,000 workers using AI across 10 countries. It describes four modes—delegation, collaboration, asking and exploration—and argues that organizations need evaluation processes as agent execution grows. The framework is qualitative, not a measured ranking of occupations or workers. Microsoft WorkLab’s 2026 Work Trend Index
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What skills matter when agents can execute more work?
The studies do not establish a single required skill list for every developer. Their findings point instead to capabilities that help people direct work and judge whether an agent’s output is fit for use:
- Specify outcomes and constraints: turn a broad request into observable acceptance criteria, including what must not change.
- Provide useful context: explain relevant architecture, conventions, dependencies and edge cases rather than treating a prompt as a substitute for domain knowledge.
- Review and test: inspect changes, run appropriate tests and look for plausible errors that may pass a superficial review.
- Supervise execution: decide where approval is needed, notice when an agent is off track and know when to stop or roll back.
- Keep technical judgment current: maintain enough understanding to assess the proposed approach, troubleshoot failures and explain the resulting system.
- Learn with the team: share effective practices and align tool use with team policies and the development lifecycle.
These are practical implications of work that includes more delegation and oversight; they are not evidence that foundational programming knowledge has become unnecessary.
How should a team evaluate an AI coding agent?
Do not judge an agent by a claim that it “writes code faster.” Compare it on the work your team actually needs done, and make the human review point explicit.
- Name the task: distinguish routine completion from debugging unfamiliar code, planning a change, testing, deployment or maintenance.
- Set the autonomy boundary: decide whether the system may only suggest, may execute a bounded task, or may take multiple steps—and where a person must approve an action.
- Define verification: specify relevant tests, review requirements, observable signs of successful completion and rollback controls.
- Check context and expertise: ask whether the developer using the tool understands the problem well enough to spot a plausible but incorrect answer.
- Assess the team setting: account for training, peer learning, organizational policy and integration with the development lifecycle.
- Read the evidence type carefully: separate controlled experiments from surveys, interviews and telemetry tied to one product.
JetBrains’ survey also reports adoption across named tools, but adoption shares are not product recommendations. The survey describes its tool-adoption findings; choosing a system still requires evaluating it against the team’s tasks, controls and standards.
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.
