From autocomplete to four phases
The framework is described in a GitHub developer guide published in November 2026. The four stages are Plan, Build, Verify and Observe, and they cover the whole software lifecycle rather than the moment a suggestion appears in the editor.
Planning means turning a requirement into something an agent can act on. Building is the part everyone already knows from Copilot autocomplete, only wider: multiple agents editing files, opening branches, running tests. Verifying is review, testing and governance. Observing is production, where the triage of incidents begins.
The tooling underneath has been growing across September and October 2026. GitHub introduced the Copilot SDK for agentic workflows, a Copilot CLI and app, a shared runtime, and a research preview called Project HydraFusion. In October 2026 GitHub added scoped firewall compatibility checks and secure network boundary controls for provider-hosted web tools inside GitHub Agentic Workflows. That is plumbing for agents that reach outside the repository, and it suggests where the real work sits: permissions, not prose generation.
The commit volume and the review queue
GitHub reports more than 1.2 billion commits in a single week during October 2026.
The same Faros AI 2026 report points the other way. Its 22,000-developer telemetry, the same basis as the 441% figure, also shows median time to first review up 156.6% and time actively spent reviewing up 199.6%. Writing code has become cheap. Reading it has not.
What 441% means in minutes
The percentage is abstract until you convert it. A review that took 30 minutes before high AI adoption takes roughly 2 hours 43 minutes at a 441.5% increase, assuming the same starting point. Twenty minutes of active review work becomes about an hour at 199.6%. This is arithmetic on the published figures, not a measurement of any particular team, but they explain why review queues feel different.
On our own RTX 5060 Ti rig in AI Arena, raw generation speed has never been the step that delays a release. The checking is. That matches what the Faros numbers describe, and it is why a framework with a dedicated Verify phase makes sense as documentation of a problem rather than a fix for it.
HydraFusion and the cost claim
GitHub positions Project HydraFusion, still a research preview, as quality-per-inference-cost orchestration. That is a research direction, not a shipping feature, and no date for general availability has been published. The direction is clear enough: GitHub is optimising the cost of running agents, not the cost of reviewing their output.
The European part is Article 50, not Annex III
Since 2 August 2026, transparency obligations under Article 50 of the EU AI Act are mandatory for deployers in the EU. Chatbot interactions must be disclosed, and synthetic content must carry machine-readable marking. On the same date the European Commission and the AI Office gained full operational enforcement powers under Article 101, with administrative fines of up to €15 million or 3% of global annual turnover.
For a European team running coding agents, that changes the paperwork around the pipeline. Agent-authored commits and agent-generated text inside documentation need to be attributable. The practical consequence is that logging and provenance become part of the Build and Observe phases, not an afterthought.
High-risk obligations are a separate story. The Digital Omnibus on AI, Regulation (EU) 2026/1744, deferred the Annex III compliance deadline to 2 December 2027, and to 2 August 2028 for AI embedded in regulated products under Annex I. Most coding assistants sit outside Annex III. Teams building AI into medical or machinery products fall under Annex I and get the longer runway.
What the guide does not settle
The human stays in the loop by design. Review, testing, governance and architectural decisions remain with people, and the framework says so without quantifying how much human attention each phase needs. There is no threshold for when a team should stop letting agents open pull requests, and no capacity model for review.
The congestion is described, not solved. A four-phase vocabulary helps a team explain to itself where the hours went after adoption rose. Reading 1.2 billion weekly commits still has to happen somewhere.
Does the four-phase model remove the need for code review?
No. Review sits inside the Verify phase and stays with human reviewers. The Faros AI telemetry in the same material shows review time rising rather than falling, which is the opposite of what removing review would look like.
Does the guide say how much human review is enough?
It does not. No ratio of reviewers to agent-generated pull requests is given, and no threshold for pausing agent activity. Teams have to set that themselves.
Are coding agents covered by the EU AI Act's high-risk rules?
Most standalone coding assistants are not Annex III systems. They fall under the Article 50 transparency duties that became enforceable on 2 August 2026. AI embedded in regulated products under Annex I has a compliance deadline of 2 August 2028.
