AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Nov 6, 2026

GitHub published a developer guide arguing that AI-assisted coding has shifted the bottleneck from writing code to planning, reviewing and operating it. The post reports that in October the platform recorded more than 1.2 billion commits in a single week, over 400% higher than a year earlier and more than GitHub received in all of 2025. It pairs that figure with a Faros AI finding that pull requests spent 441% longer in review when teams moved from low to high AI adoption. The guide was also presented live at GitHub Universe 2026 in a session titled "Agentic Engineering: Past the Pilot."

The argument runs through four stages: Plan, Build, Verify and Observe. In each, GitHub positions Copilot less as a generation engine and more as a coordinator that gathers context, runs work in the background, and clears routine checks so humans spend attention on decisions. Typical planning pain is described concretely: an API change whose design lives in one tool, a compatibility requirement buried in chat, and an old pull request explaining why the current implementation works as it does. GitHub's answer is to pull that scattered material into one place, convert agreed requirements into Issues with acceptance criteria, and use Plan Mode to compare approaches before any code is written.

Triage gets automation too. GitHub Agentic Workflows can label issues and estimate complexity, with low-complexity items assigned directly to Copilot while harder work gets human planning. Collaboration is meant to happen where the team already is, through Copilot in Slack or Teams, with context drawn from connected tools.

For the Build stage, the features center on continuity and isolation. Session continuity lets a developer resume the same Copilot session across clients and devices. Tasks execute in cloud sandboxes rather than on the local machine, so a session survives a closed laptop. The Copilot app gives each session an isolated copy of the repository, so an implementation task and a documentation update do not share a half-finished working tree. Repository instructions capture conventions and validation commands, and shared memory stores facts Copilot reuses across a repository. A feature called Copilot Dreaming reviews eligible sessions and pull request feedback to keep those memories current and to propose instruction or skill updates. Copilot Fusion selects models and reasoning depth per task, and can bring in additional models to review a solution, with the effort level tuned for quality, cost and latency. Enterprise managed settings carry company policy and configured tools across Copilot surfaces.

Verify is where the 441% figure gets addressed directly. The guide recommends stacked pull requests so dependent work lands as smaller reviewable pieces, and GitHub Code Quality checks changes against organizational standards before a human opens the diff. Copilot code review adjusts analysis depth to the size of the change, either user-selected or chosen automatically. AI Scan extends vulnerability detection to languages and files beyond CodeQL's coverage, and AI secret detection targets unstructured secrets such as passwords that pattern matching misses. License compliance checks dependencies against company rules, and the GitHub package firewall, in public preview for npm, can block downloads that violate configured security policies. Agentic Autofix proposes repairs for security findings, while Agent Merge works through failing checks, review comments and merge conflicts; if configured, Copilot merges once verification passes.

Observe covers post-release investigation. Using connectors that include Fabric, Copilot can help determine which requests or customer groups a problem affects, then draft an issue containing supporting evidence, acceptance criteria and a regression test. Reviewing Dreaming's suggested improvements is meant to feed lessons back into memories, skills and review guidance so the next planning cycle starts from what production revealed.

My reading is that the interesting claim is organizational rather than technical. If generation capacity rises sharply, review bandwidth becomes the scarce resource, and most of these features are really queue-management and context-management tools. Stacked PRs, tiered review depth and pre-merge quality gates all aim at the same constraint. That has trade-offs worth watching: automated autofix and auto-merge shift trust onto configured checks, so policy accuracy matters more than it did when a human caught everything; shared memory and Dreaming introduce a maintenance surface that teams must curate or risk stale conventions propagating into new work; and running agents in cloud sandboxes raises questions about where code and credentials live. The package firewall being npm-only in preview also limits how much supply-chain protection teams get today.

Why it matters: Teams adopting coding agents heavily are likely to see review queues, not authoring, become the limiting factor, so tooling that shrinks and pre-screens changes may matter more than raw generation speed. Developers and platform teams should watch the verification layer, especially Code Quality, AI Scan, Autofix and Agent Merge, since that is where trust in autonomous work is earned or lost. Product teams planning rollouts may want to pilot the Plan and Verify stages first, given the review-time evidence GitHub cites, though whether these features close the gap is not yet demonstrated.

Read at GitHub Blog · AI & ML

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Writing code got faster. The work around it didn't. See how you can plan, write code, review changes, and investigate production issues with GitHub Copilot, and what work still needs your judgment. The post Beyond code generation: A developer’s guide to agentic engineering appeared first on The GitHub Blog .