This report sets out eight trends Anthropic predicts will define agentic coding in 2026, grouped into three categories: foundation trends that reshape how development work happens, capability trends covering what agents can accomplish, and impact trends affecting business outcomes and organisational structures. The document is explicit about its own status — the predictions are offered as “a framework for thinking about the year ahead” reflecting what the company sees with customers today, not as certainties.
Its organising observation is a tension the report returns to repeatedly. Research from Anthropic’s Societal Impacts team is reported as finding that developers use AI in roughly 60% of their work while being able to “fully delegate” only 0–20% of tasks, which the report reads not as a contradiction but as evidence that effective use is collaborative: it requires set-up, prompting, active supervision, validation, and human judgement, especially for high-stakes work. The report treats this as the reason the human role stays central even as capability grows, and the shift it describes is from writing code to reviewing, directing, and validating it. This pattern is developed under Collaboration Paradox.
The report also argues the gap between early adopters and late movers is widening, and that the organisations best positioned are those that work out how to scale human oversight without creating bottlenecks.
Details
The eight trends, as the report names them:
- The software development lifecycle changes dramatically. Tactical writing, debugging, and maintenance shift to AI while engineers move to architecture, system design, and strategic decisions; the engineering role becomes orchestrating agents and evaluating their output; onboarding to a new codebase collapses from weeks to hours, enabling dynamic “surge” staffing.
- Single agents evolve into coordinated teams. Organisations adopt multi-agent workflows that parallelise reasoning across separate context windows, requiring new skills in task decomposition, agent specialisation, and coordination protocols, plus tooling that shows concurrent agent sessions and version control that handles simultaneous agent-generated contributions.
- Long-running agents build complete systems. Task horizons expand from minutes to days or weeks, with agents planning and refining across dozens of sessions; the report argues this changes project economics, making formerly non-viable projects feasible and letting accumulated technical debt be worked through systematically. See Long-running Agent.
- Human oversight scales through intelligent collaboration. Agents learn when to ask for help rather than blindly attempting every task; AI agents review large-scale AI-generated output for security and architectural consistency; human oversight shifts from reviewing everything to reviewing what matters.
- Agentic coding expands to new surfaces and users. Support extends to less-common and legacy languages such as COBOL and Fortran, and new form factors open agentic coding to non-traditional developers in cybersecurity, operations, design, and data science.
- Productivity gains reshape software development economics. The report names three compounding multipliers — agent capabilities, orchestration improvements, and better use of human experience — and reports that internal research found a net decrease in time per task category alongside a much larger net increase in output volume, concluding that the gain comes primarily through greater output rather than doing the same work faster. It also reports that about 27% of AI-assisted work consists of tasks that would not otherwise have been done at all.
- Non-technical use cases expand across organisations. Teams in sales, marketing, legal, and operations automate workflows without engineering intervention, and domain experts implement solutions directly rather than filing a ticket and waiting.
- Agentic coding improves security defences — but also offensive uses. The report frames this as dual-use: any engineer can perform security reviews and hardening that previously needed specialist expertise, while the same capabilities help attackers scale, making it more important to build security in from the start. It also predicts agentic cyber defence systems responding at machine speed.
For organisations planning the year, the report names four priorities: mastering multi-agent coordination to handle complexity single-agent systems cannot address; scaling human-agent oversight through AI-automated review systems that focus human attention where it matters most; extending agentic coding beyond engineering to empower domain experts across departments; and embedding security architecture as a part of agentic system design from the earliest stages.
Notes
The report supports most trends with a named customer example, alongside aggregate internal research in the trends on oversight and productivity. Among them: Rakuten engineers completing in seven hours of autonomous agent work a vLLM implementation task in a 12.5-million-line codebase, reported as reaching 99.9% numerical accuracy against the reference method; Fountain, a workforce-management platform, using a central orchestration agent over specialised sub-agents; CRED, a fintech platform, reporting doubled execution speed; TELUS reporting over 13,000 custom AI solutions and engineering code shipped 30 percent faster; and Zapier reporting 89 percent AI adoption across the organisation. These are the report’s own accounts of its customers’ results, and it does not describe how they were measured.
The report is a vendor publication about the vendor’s own product category, and it says so in its framing — the examples are Anthropic customers and the internal research cited is Anthropic’s own.