Completed Work, Not More Code
At GitHub Universe, the message was clear: the job is shipping finished software with agents, not producing more lines. Here is what that changes about review, accountability, and what done means.
4 min read
At GitHub Universe on October 28, 2026, Jay Parikh put the shift in plain language. In a couple of years, developers have gone from writing every line of code, to working with assistants, to orchestrating fleets of agents. The goal, he said, has not changed: ship completed software work. A feature, a bug fix, a security fix. Completed work, not more code, is the output that matters.
That framing is useful because it cuts through a lot of noise. More pull requests are not progress if nobody can review them. Faster generation is not delivery if the change never lands safely. Agents change the volume and shape of the work. They do not change who is accountable for what ships.
What GitHub is building around that idea#
GitHub used Universe to describe an agentic delivery system with three aims: velocity, intelligence, and quality. The products around that idea are less interesting as features than as a statement of the workflow.
The GitHub Copilot app is meant to take work from issue to merge in one surface. Copilot projects run agent work across multiple repositories in parallel managed sessions. Session continuity lets the same cloud agent session move between the app, VS Code, the CLI, and the web. Automations can respond to events, run on a schedule, and propose improvements from their own outcomes.
Copilot Fusion sits underneath a lot of that. It routes and orchestrates models by task instead of asking a developer to pick a model every time. On the Artificial Analysis Coding Agent Index, Fusion Max scored within two points of Opus 5.5 Max at about half the cost and about 40% less time per task.
The same day, VS Code 1.144 shipped an Agent Host that keeps sessions running when the client closes, an Agents window for managing work across environments, canvases for richer interaction, and experimental computer use so agents can drive desktop apps. The editor is becoming a control surface for ongoing agent work, not only a place to type.
The metric that already moved#
The volume change is not theoretical. In October 2026, CNBC reported that GitHub said one in three pull requests now involve an AI agent. A year earlier that figure was fewer than one in ten. GitHub's own comment in that reporting was blunt: if the pace holds, most code pushed to GitHub could be written by an agent within a couple of years, and much of it may never be fully read by a human.
That is the real product problem. Not whether agents can write code. Whether teams can keep quality, security, and accountability as the share of agent-authored changes climbs.
What done has to mean#
If completed work is the unit of value, a few definitions have to get sharper.
A task is not done when the agent finishes generating. It is done when the change is reviewed against a standard, tests and checks pass, and a human who owns the outcome is willing to merge it.
Review has to scale with volume. Copilot code review Max was introduced because more code means more pull requests. Deeper review for high-severity issues is one answer. Human judgment on architecture, product fit, and risk is still another.
Automations should learn from outcomes. GitHub's automations can propose improvements from completed runs. That is the right direction. A loop that only produces more draft diffs is a treadmill. A loop that records what failed and changes the next run is a system.
Accountability stays with people. Agents implement and verify. Developers set direction and remain accountable for what ships. That sentence from Universe is the one I would keep on a whiteboard.
A practical way to use the news#
I would treat this wave of announcements as a prompt to rewrite the definition of done for agent work, not as a shopping list.
Ask, for each kind of change:
- What evidence is required before merge?
- Who can approve what the agent may do without asking?
- What gets recorded when a run fails so the next run is better?
- How do we keep humans reading the changes that matter most when agents are producing the rest?
Completed work is a better north star than lines of code. The hard part is making that concrete in the merge gate, the review path, and the learning loop around both.
Related
- Leadership
The Job Didn't Disappear. It Moved Up the Stack.
Software job postings are rising even as agents write more of the code. The rebound is concentrated in senior and AI-oriented roles. Entry-level paths are thinner. That changes how teams hire and how they grow people.
Careers · Leadership · Agents · AI3 min read
- Engineering
Nothing Goes Live Because It Exists
In my agent rig, prompts, routes, skills and even fine-tuned models all start in shadow. Promotion is a separate, evidence-backed step that a person takes. Here's the lifecycle, and why it matters more as the number of moving parts grows.
Agents · AI · Local LLMs · Delivery6 min read
- Engineering
The Art of the Rock Tumbler: Polishing Loops for Code and Agents
A rock tumbler doesn't make a stone smooth in one pass. It uses coarse grit, then finer grit, then polish, and checks between stages. That's how I build agent loops in Starship: bounded retries, deterministic gates, and lessons that are only allowed to hurt once.
Agents · AI · Architecture · Delivery7 min read