Coding assistants started as suggestion engines inside the editor. Agentic tools plan multi-step work, use a shell and other tools, and iterate on the results, so they can be handed a task rather than a line to complete. The impact is therefore felt well beyond writing code: in how requirements are written, how changes are reviewed, how tests are produced and how releases are gated.
01From Assistant to Agent
The early evidence was about assistants. In a controlled experiment reported by Peng and colleagues in 2023, developers with GitHub Copilot completed a specific task, writing an HTTP server in JavaScript, 55.8% faster than a control group. That is a clear result for a narrow, well-defined task. Agents widen the scope: the unit of delegation becomes a ticket or a bug report, and the work includes reading the repository, running commands and fixing its own failures. The Agent Harnesses and AI Coding Agents pages on this site cover the tooling.
Wider scope means the effects spread across the lifecycle, and the bottleneck moves. When producing code gets cheaper, understanding the problem and verifying the result take up a larger share of the effort.
02Requirements and Planning
Agents are useful for turning rough input into structured drafts: user stories, acceptance criteria, edge cases and open questions, produced from tickets, call notes or existing documentation. They can also check a new requirement against existing behaviour in the codebase and flag conflicts.
- What improves. Faster first drafts, more consistent acceptance criteria, and a wider list of edge cases to discuss.
- What gets harder. A vague requirement now produces a confident, plausible implementation of the wrong thing, and does so quickly. Precise intent, written down, becomes the main input to the whole process.
- Practice. Treat the specification as the artefact that is reviewed first. A practical option is to keep it as a versioned file in the repository so that agents and humans work from the same text.
03Architecture and Design
Agents can explore options, summarise how an unfamiliar module works, draft design documents and sketch interfaces. They are less reliable at judgments that depend on context not visible in the repository: the organisation's constraints, long-term roadmap, regulatory position or operating budget.
- What improves. Faster discovery in legacy code, quicker comparison of alternatives, and drafts of architecture decision records.
- What gets harder. Local solutions accumulate. An agent optimising for the task in front of it can duplicate logic or bypass an agreed pattern, so architectural consistency has to be enforced through written conventions, linters and review rather than assumed.
04Implementation
This is where agents are most visible. They scaffold services, implement features, migrate frameworks, upgrade dependencies, fix bugs and write boilerplate. Work that is repetitive, well specified and checkable by tests is the best fit.
The measured picture is mixed, and it is changing. In METR's 2025 randomised trial, 16 experienced open-source developers completed 246 real tasks in repositories they knew well. With early-2025 tools they took 19% longer when AI was allowed, even though they had forecast a 24% speedup and afterwards believed they had been about 20% faster. METR has since marked those results as out of date. Its February 2026 update, with 57 developers, 143 repositories and more than 800 tasks, estimated changes in task time of -18% for returning developers and -4% for new ones (negative means faster), both with wide confidence intervals that include zero. METR says those estimates are unreliable because many developers declined to work without AI and withheld tasks, and believes the true speedup is likely larger than the estimates show, without quantifying it.
What to take from the studies
Perceived speed and measured speed can differ, results depend heavily on the tool generation, the developer's familiarity with the codebase and the type of task, and no single study settles the question. Measure your own cycle time, rework and defect rates instead of relying on how fast the work feels.
05Testing and Quality Assurance
Agents generate unit tests, propose property-based tests, reproduce bug reports as failing tests and triage flaky suites. Because an agent can run the tests and iterate, a strong test suite becomes the specification the agent works against.
- What improves. Coverage of routine cases, faster bug reproduction, and cheaper regression tests for legacy code.
- What gets harder. Tests written by the same agent that wrote the code can share its misunderstandings. A passing suite is only as meaningful as its assertions, so review the tests as carefully as the code, and keep independent checks such as integration tests, mutation testing and static analysis.
06Code Review and Security
More generated code means more to review. Agents can do a first pass: summarising a change, flagging risky patterns, checking style and spotting missing tests. But the human reviewer's job grows more important, not less, since the author may not have written or fully read the code being submitted.
Developer sentiment reflects this. In the 2025 Stack Overflow Developer Survey, 45.7% of respondents said they distrust the accuracy of AI output and 32.7% trust it; 66% named "AI solutions that are almost right, but not quite" as a frustration. Almost-right code is the hardest kind to review, because it looks correct.
- Keep pull requests small and ask agents to explain their changes and the checks they ran.
- Run secret scanning, dependency and licence checks, and static security analysis on every change regardless of who or what wrote it.
- Give agents least-privilege credentials, sandboxed environments and no direct path to production. The Guardrails for Agentic AI page covers the controls.
07Release, Deployment and Operations
Agents help with pipeline configuration, infrastructure-as-code drafts, release notes, incident summaries and first-line diagnosis of alerts. In operations they can gather logs and metrics, propose a rollback or a fix, and prepare the post-incident write-up. Actions that change production should stay behind approval gates and be reversible.
The 2025 DORA report, based on nearly 5,000 technology professionals, gives the most useful framing here. Google's announcement says AI adoption now has a positive relationship with delivery throughput but still a negative relationship with delivery stability, which the authors attribute to faster change exposing weaknesses downstream. Without strong automated testing, mature version control and fast feedback loops, more change volume leads to more instability. The report's headline is that AI "doesn't fix a team; it amplifies what's already there".
08Phase-by-Phase Summary
| Phase | What agents contribute | New risk | Control to add |
|---|---|---|---|
| Requirements | Draft stories, criteria, edge cases | Fast delivery of the wrong thing | Reviewed, versioned specifications |
| Design | Options, summaries, decision records | Inconsistent local solutions | Written conventions, architecture review |
| Build | Features, migrations, bug fixes | Unread or misunderstood code | Small changes, explanations, tests as gate |
| Test | Test generation, bug reproduction | Tests that share the code's blind spots | Independent checks, mutation testing |
| Review | First-pass review, summaries | Reviewer overload, almost-right code | Human approval, security scanning |
| Release | Pipeline and infrastructure drafts | Higher change volume, less stability | Progressive delivery, rollback plans |
| Operate | Diagnosis, runbooks, incident notes | Autonomous actions in production | Approval gates, audit logs, least privilege |
09Roles, Skills and Measurement
The developer's role shifts toward specifying, delegating, reviewing and integrating. Reading code critically, writing precise intent and designing for testability matter more. Junior developers still need to learn how systems work, so teams should keep deliberate practice in the process rather than let delegation replace it; this is a judgment call rather than a measured finding.
Measure outcomes, not activity. Lines or pull requests produced rise easily with agents; lead time, change failure rate, time to restore service and rework say whether delivery is actually better. These are the DORA delivery metrics, and tracking them before and after adoption gives a team its own evidence.
10Putting It Together
Agentic AI speeds up parts of the lifecycle and moves the constraint to specification, verification and operations. The evidence so far supports modest optimism with caution: throughput gains are reported, stability and trust remain concerns, and the research itself is still catching up with the tools. Teams that already have clear requirements, good tests, small changes, fast pipelines and solid review practices are best placed to benefit, which is the amplifier point from DORA. Start with well-bounded tasks, keep humans on the gates that matter, and measure the result.
Sources and further reading:
- Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (METR) — the randomised trial, now marked out of date by METR.
- METR's February 2026 update on developer productivity — the late-2025 estimates and why METR considers them unreliable.
- Announcing the 2025 DORA Report (Google Cloud) — AI as an amplifier, throughput and stability findings.
- Stack Overflow Developer Survey 2025: AI — usage, trust and frustration figures.
- The Impact of AI on Developer Productivity: Evidence from GitHub Copilot — the 55.8% faster result on a defined task.
- On this site: AI Coding Agents, Agent Harnesses, Agentic AI, Guardrails for Agentic AI, Open-Source Benchmarks for AI Coding Agents.