Developers stop writing code and start orchestrating agents

The scale of agent-started work

Developer activity on GitHub reached record levels in 2026, with developers pushing code 17.9 billion times in the past 12 months and commits across all repository types climbing to 34.3 billion in 2026. GitHub now serves more than 225 million developers, and GitHub Actions ran a job 1,440 times every second in 2026, with 5.8 billion push protection scans run in the same year. The volume changes the job: review and orchestration now matter as much as writing the first draft. m:…4430

As of September 2026, more than 40 percent of pull requests on GitHub are started with an agent, and agent-started pull requests change multiple files about 87 percent of the time. Agent work is not small, single-file edits. It arrives as multi-file changes that need review across a whole change set. m:…4430

TypeScript, Python and JavaScript are the top three languages on GitHub by contributor count, and Rust grew fastest by percentage among the top languages in 2026. Repositories serving mining, construction, health care and agriculture have grown several times over since 2022, which suggests agent-assisted development is spreading into industries that were not the early adopters of developer tooling. m:…4430

Orchestration becomes the product

At GitHub Universe 2026, GitHub presented its vision of an agentic software delivery system, with the GitHub Copilot app connecting agents and tools across projects so developers can direct work without manual handoffs. Copilot projects allow users to run agent work across multiple repositories in fully managed, parallel sessions, and session continuity allows cloud agent sessions to persist across the GitHub app, VS Code, the command line interface and the web. m:…5478

In September 2026, 61 percent of GitHub Copilot users selected more than one model for their sessions, and Copilot Fusion is generally available to provide model orchestration. Fusion evaluates tasks across reasoning, code generation, debugging and tool use to choose the best model for execution, and GitHub reports that Fusion Max scored within two points of Opus 5.5 Max at half the cost and 40 percent less time per task. Copilot automations allow users to create AI-powered workflows that learn from outcomes to propose improvements, and Copilot code review Max posted the highest F1 score on ReviewBench for high-severity issues. m:…4430 m:…5478

GitHub Enterprise customers can use all Copilot features without a separate subscription by paying for AI credits, and GitHub added 4 million CPU cores, 160 petabytes of high-speed storage and 50 terabits per second of network capacity this year. The pricing shift to credits points the same way as the infrastructure build: agent work is expected to be metered and parallel, not occasional. m:…5478

How you run the agent changes the bill

Pydantic AI internal benchmarking of code mode, where models write and execute code loops instead of calling tools one by one, found that code mode roughly halved cost and wall time for long tasks across eight models. Success rates for long tasks of 200 items dropped to 62 to 79 percent without code mode, compared with 92 to 100 percent with it. Code mode reduces the volume of tokens processed in each request rather than just reducing round trips, and it allows models to design workflows on the fly by writing code that acts as a graph. m:…8278

The same benchmarking found a trade-off at small scale. At 5 items, code mode cost 21 to 38 percent more than standard tool calling on three of five tasks, and code mode guidance adds a flat fee of approximately 2,600 characters per request. Models like gpt-6.1-sol and Claude models already perform batching of tool calls without code mode, and Pydantic notes that speculative execution showed higher costs and lower success in initial tests, while code mode failures can lead to repeated writes if a snippet is rerun after a partial failure. The lesson is to match the orchestration to the task size rather than treating one mode as always cheaper. m:…8278

The interpreter used in Pydantic AI Harness, Monty, starts with no host access and enforces a 30-second execution limit, which is a reminder that letting a model write and run code needs a sandbox with real limits. Replit makes the same point at platform level: it uses microVMs to provide isolated execution environments for AI agents, its credential broker retrieves secrets from a vault and adds them to requests only after authorization, and its Agent Habitat framework provides governance tools including audit logs and role-based permissions. m:…8278 m:…7843

Replit runs AI red teaming around the clock, offers Bring Your Own Cloud so companies can host agent harnesses in their own cloud tenant, and is developing an agent-action policy system with organization-wide limits. Replit also reports that internal code output for a consistent cohort of engineers grew 2.9 times between January and June. That company-reported growth figure should be read alongside the isolation detail, because output growth without isolation is a liability, not a win. m:…7843

What changes for developers

This suggests the developer job is becoming a review and orchestration job. When more than 40 percent of pull requests start with an agent, when most Copilot users already pick more than one model, and when the orchestration choice can halve the cost of a long task or raise it on a short one, the skill that matters is judging a multi-file change, picking the right model and mode for the task, and keeping agent execution inside a sandbox with audit logs. Writing the first draft is no longer the scarce step. m:…4430 m:…8278