The Developer Is the Next Migration: From Writing Code to Orchestrating Agents

Anthropic’s 2026 Agentic Coding Trends Report states its through-line up front: software development is “shifting from an activity centered on writing code to an activity grounded in orchestrating agents that write code,” while “maintaining the human judgment, oversight, and collaboration that ensures quality outcomes.” The second clause matters as much as the first. The shift is not from human to no-human. It is from one kind of work to another.

The report names the new role directly. Its first trend is role transformation, from implementer to orchestrator. The value of an engineer’s contribution, it argues, moves to “system architecture design, agent coordination, quality evaluation, and strategic problem decomposition.” That is a claim about where senior engineering time goes, not a claim that engineering time goes away.

The most important number in the report undercuts the hype. Anthropic’s Societal Impacts team found that developers use AI in roughly 60% of their work but can fully delegate only 0 to 20% of tasks. The report calls this a collaboration paradox. The gap between those two numbers is where the role shift actually happens. What changes is not whether a human is involved but what the human is doing: reviewing, directing, and validating generated code instead of typing it.

I should say what this report is before leaning on any of its numbers. Anthropic sells Claude Code, the coding agent the report describes. Every customer statistic in it is first-party and unaudited: TELUS shipping engineering code 30% faster and saving more than 500,000 hours, Rakuten’s engineers getting Claude Code to implement a vLLM activation-vector method across a 12.5-million-line codebase in seven hours at 99.9% numerical accuracy, Zapier reaching 89% adoption with more than 800 internal agents. The eight trends are framed as predictions, not measurements. The foreword says they reflect what the company is seeing with customers today, not certainties about tomorrow. I am treating this as a vendor’s account of its own product, and nothing that follows rests on any single one of those numbers.

The mechanism is delegation, not suggestion

The shift has a mechanism worth separating from the vendor’s framing. Firecrawl’s trends list calls it “delegation over suggestion.” An IDE assistant offers a suggestion and waits for a human to approve each change. A terminal agent, Claude Code, Cursor, OpenCode, Windsurf, runs on its own for hours, changes files, runs shell commands to check its work, and commits. The difference is not just autonomy. It is where the human’s attention goes. Firecrawl puts a number on the economics: a CLI call averages around 200 tokens, against 32,000 to 82,000 for the equivalent MCP operation. That token math is what pushes production pipelines toward the terminal.

The independent confirmation comes from Vela Partners’ scan of open-source GitHub activity. Terminal-based AI coding agents is its second-largest category by stars, 561,600 across Claude Code, OpenCode, Aider, Cline, and OpenHands, up 74,600 in 90 days. Vela’s conclusion is blunt: “developer tools are the new distribution moat.” The caveat sits in Vela’s own methodology. It clustered 854 repositories using Gemini 3.1 Pro, and it notes itself that “stars are noisy.” Stars measure attention, not production adoption. I am citing the number as evidence of where attention is going, not as proof of anything more.

What this does not mean

There is a version of this story that overreaches, and it is worth not writing. The role shift does not mean the governance layer disappears. Firecrawl’s own data shows MCP usage grew about 35% in a month, and Zuplo’s State of MCP survey found 72% of adopters expect to increase use, with 54% expecting it to persist or become a standard. The CLI wins on token efficiency. MCP wins on authentication, multi-tenancy, and governance. Both are growing because they solve different problems. I covered the MCP side in the rise of MCP governance.

The infrastructure thesis is also already on record here. The model isn’t the bottleneck anymore made the case that the harness around the model is the product, and it carried OpenAI’s counterpoint that the product is still model-limited. I am not re-arguing either side. What is new in this report is the claim about the developer, not the claim about the stack.

The closest thing I have written to this is the verification gap, which argued from Dan Luu’s experiment that agents cannot check their own work even when told how. The 0 to 20 percent delegation figure is the same gap seen from the vendor’s side. Anthropic’s own research concedes that only a small slice of a developer’s work can be handed off entirely. That is not a flaw in the pitch. It is the part that makes the pitch believable.

What I find useful, and what I do not

The useful part is that the role shift gives senior engineers a concrete job description. Architecture, coordination, evaluation, decomposition. Those were always the hard parts. The report is arguing that the easy part, typing the code, is what gets automated first, and that reads right to me.

The slippery part is the productivity math. The report notes that engineers report “a net decrease in time spent per task category, but a much larger net increase in output volume,” and that roughly 27% of AI-assisted work is tasks that “wouldn’t have been done otherwise.” Output volume is easy to measure and easy to inflate. Whether that volume is code that should exist is the question the report does not answer.

It is also worth saying that “orchestrator” is a promotion nobody gets handed. Most working developers are not going to spend their days on architecture while agents do the typing. They are going to spend their days reading diffs, checking whether the agent’s change matches the ticket, and deciding whether the extra 27% is worth the review time. That is still the developer’s job. It is a less glamorous version of it than the report’s framing suggests.

Bottom line

The report’s through-line is right, and it is not another argument about the stack. Software development is shifting from writing code to orchestrating the agents that write it. The number that matters is the gap between the 60% of work that uses AI and the 0 to 20% that can be fully delegated. That gap is where the developer’s next job actually lives, in reviewing, directing, and validating rather than implementing. The customer stats, the token math, and the star counts are supporting evidence, and some of them are vendor marketing. The role shift does not depend on them.

Sources

1. 2026 Agentic Coding Trends Report — Anthropic
2. Top 15 Agentic AI Trends to Watch in 2026 — Firecrawl
3. 19 Open-Source AI Infrastructure Trends 2026 — Vela Partners