Beyond code generation: A developer’s guide to agentic engineering

This October, GitHub saw more than 1.2 billion commits in one week—over 400% more than a year earlier. This is more commits than GitHub received in all of 2025. Chances are, you’re generating more code than you were a year ago. But ask yourself: do you feel 400% more productive than you were last year?

Building software takes more than a good prompt. As AI speeds up implementation, software developers are feeling more pressure in other places: a Faros AI study found that pull requests spent 441% longer in review when teams went from low to high AI adoption.

As a software developer, you gather requirements and steer implementation, review changes and evaluate test results, then investigate what happens in production. Agents can help you work more efficiently across all of these stages: Plan, Build, Verify, and Observe. We’ll walk through what an effective agentic engineering workflow looks like on GitHub.

Plan: Get everyone and every agent on the same page

See if this sounds familiar. You’re planning an API change. The design is in one tool, a compatibility requirement is buried in Slack, and an old pull request explains why the current implementation works this way. Before you can write a useful plan, you need to connect those details, compare approaches, and agree on how existing clients will keep working. What looked like a straightforward planning task has become hours of searching, cross-checking, and chasing answers. You’re exhausted before implementation even begins.

You can use Copilot to pull that scattered context into one place, work through the tradeoffs, and turn the decisions into an implementation plan.

• Plan together where your team already works. Use Copilot in Slack or Teams to collaborate with everyone involved—from product managers and designers to analysts and business stakeholders. Bring in context from connected tools so your team can check assumptions, contribute ideas, and refine requirements together.
• Turn agreed-upon requirements into actionable work. Ask Copilot to create GitHub Issues with the context and acceptance criteria needed for implementation, breaking out tasks that can be handled separately.
• Keep routine triage from consuming planning time. Use GitHub Agentic Workflows to label your issues and estimate task complexity. Assign low-complexity tasks directly to Copilot, then spend your time focusing on harder decisions.
• Give complex work the planning it needs. Use Plan Mode in your preferred Copilot surface to compare approaches and refine an implementation plan. Review and approve the plan before handing it off to Copilot to begin implementation.

What took days of searching and context gathering now produces a reviewable plan, explicit acceptance criteria, and scoped issues, with your time spent on the decisions instead of the retrieval. Then, once you settle on an approach, you’ve already set yourself and your agents up to implement your plan.

Build: Move more than one task forward at a time

You’re ready to turn that plan into code. But life has other ideas. Your agent is halfway through a task, your next meeting starts in five minutes, and your dog has been staring at the door for twenty. Then someone in the meeting points out an edge case the implementation needs to handle. Everyone needs something from you at once. You’d like to keep making progress, but you can’t be in three places at the same time.

You can keep your work moving without being tied to the computer where you started it.

• Continue your conversation from another device. Session continuity lets you pick up the same Copilot session across clients, so you can work wherever it’s most convenient for you.
• Let work continue after you close your laptop. Copilot runs tasks in cloud sandboxes, keeping execution off your machine and allowing you to leave and return to a session later.
• Work on separate tasks without mixing their edits. The Copilot app gives sessions isolated copies of the repository, so an implementation task and a documentation update don’t share an unfinished working tree.
• Carry project knowledge into each task. Repository instructions capture conventions and validation commands, while shared memory stores useful facts Copilot can reuse across a repository. Copilot Dreaming reviews eligible sessions and pull request feedback to keep memories current and propose instruction or skill updates. That means less time re-explaining how your team works and more consistent help from Copilot across tasks.
• Match the models to the work. Select Copilot Fusion to choose models and reasoning depth for the task, bringing in additional models to review a solution when needed. You set the effort level to balance quality, cost, and response time.
• Follow company rules wherever you work. Enterprise managed settings carry your company’s policies and configured tools across Copilot surfaces, reducing setup differences as you move between clients.

Writing code is more than shooting off a single prompt to your LLM and seeing what you get back. And with more work on your plate than ever, it can be difficult to maintain your coding standards and deliver your projects on time. Copilot helps tasks progress independently with shared project knowledge, while letting you pick up a session where you left off across devices. That gives you more room to work through implementation decisions and handle changes as they come up. And as your work finishes, you have focused changes ready to review.

Verify: Spend your attention where it counts

After spending an hour wrangling your agents to get them moving in the right direction, you think you’ve all landed in a good spot. So you open a pull request labeled “small refactor.” It changes 47 files. You wanted feedback on shared validation logic, but it’s buried among updates to every endpoint that calls it, formatting changes, and several hundred lines of generated tests. You wonder how long it will take to get an approval on this, given the endless list of pull requests on your team’s review board.

That review queue is the 441% in practice. You can shorten the wait from both ends: make your changes easier for a person to review, and let Copilot clear the routine checks and fixes before anyone opens the diff.

• Stack up smaller, more focused pull requests instead of creating a single, massive change. You and your agents can use stacked pull requests to separate dependent work into smaller pieces. This lets each PR be reviewed on its own, keeps dependencies visible, and unblocks teams by letting them build on top of your work before it merges.
• Catch routine quality issues before they consume review time. GitHub Code Quality checks changes against coding practices and your organization’s standards, catching issues before they reach a human reviewer and saving time for both authors and reviewers.
• Get review help matched to the change. Copilot code review gives automated feedback, with lighter analysis for simple changes and deeper analysis for complex ones. You can set that depth yourself or let Copilot select it for each pull request.
• Check security across more of your codebase. AI Scan finds vulnerabilities in languages and files beyond CodeQL’s coverage. AI secret detection helps identify unstructured secrets, such as passwords, that pattern matching can miss.
• Keep new dependencies within policy. License compliance checks dependency licenses against your company’s rules before merging. GitHub package firewall, currently in public preview for npm, can block downloads that violate configured security policies.
• Resolve findings and follow-up work without babysitting every pull request. Agentic Autofix proposes repairs for security issues, while Agent Merge automatically addresses failing checks, review comments, and merge conflicts—helping you keep your pull requests moving and preparing them for merge. And, if you configure it, Copilot will even merge your pull requests once your verification steps are passing.

Smaller changes with the routine work already handled means fewer rounds of back-and-forth. Your reviewers get to spend their attention on whether the code does what you intended, and what could go wrong.

Observe: Learn from what happens after you ship

Your change gets approved, the deployment checks pass, and you finally get to cross it off your list. The next morning, your on-call dashboard shows intermittent timeouts, and support reports that a few large customers are affected. Now you’re comparing logs, release history, and customer reports while someone asks whether you should roll back. “All checks passed” feels less reassuring than it did yesterday.

Bring operational evidence and engineering context together so you can decide what to do.

• Check assumptions against real behavior. With access to your observability tools and business data through connectors such as Fabric, Copilot can help investigate which requests or customer groups are affected.
• Capture the finding where someone can act on it. Ask Copilot to draft an issue with supporting evidence and acceptance criteria, including a regression test for the failure you uncovered.
• Carry useful lessons into future work. Review suggested improvements from Dreaming, which uses sessions and repository activity to help refine memories, skills, and code review guidance.

Figuring out what happened after a release can leave you bouncing between dashboards, conversations, and commit history. With Copilot helping connect that information, you can document what you found, create follow-up issues, and improve the guidance available for future tasks. You spend more time deciding what needs to change and less time retracing the investigation. And your next plan starts with what you learned from the software people are actually using.

Where to start

Take one piece of work already on your plate and look beyond the coding task. What context does it need? How will you review and test it? How will you know it worked after it ships?

You don’t have to change your whole process at once. Start with something that frustrates you most this week and put an agent on it. Then expand to other parts of your process over time. Once agents are helping across your whole workflow, you might finally feel as productive as your commit count says you are.

We covered this topic live at GitHub Universe 2026 in “Agentic Engineering: Past the Pilot.”

The post Beyond code generation: A developer’s guide to agentic engineering appeared first on The GitHub Blog.