Why Frontier AI Labs Are Obsessed with Coding Abilities

Code as a way to do the work #

I keep coming back to a hypothesis that still feels directionally correct: work that can be expressed as code can be done by large language models. When frontier labs talk about "agents," I hear a very specific capability target: turn intent into an executable plan, run it in a controlled environment, and produce outputs that can be verified.

Anthropic's MCP code-execution write-up is a clean case study. The demo flow is Google Drive → Salesforce, and it's also the shape of a lot of white-collar reality: permissions, schema mapping, formatting rules, and audit trails across systems.1Anthropic, "Code execution with MCP: building more efficient AI agents"https://www.anthropic.com/engineering/code-execution-with-mcp The key move is context efficiency. The agent discovers tools through a filesystem-like registry, loads only the interfaces it needs, and does the transformation work inside a sandboxed runtime. Anthropic reports dropping one workflow from ~150,000 tokens to ~2,000 tokens, a 98.7% reduction, by keeping large intermediate data out of the model's context and letting code handle the plumbing.

Beyond APIs #

The browser and desktop layer is where this expands from "API automation" into "office work." Many systems still force you through a UI, so the agent needs to see pixels, click, type, and maintain state across screens. OpenAI's Computer-Using Agent results give a useful snapshot: 58.1% success on WebArena and 87% on WebVoyager for web tasks, plus 38.1% on OSWorld for full computer-use tasks.2OpenAI, "Computer-Using Agent" (WebArena, WebVoyager, and OSWorld results)https://openai.com/index/computer-using-agent/ For context, human performance on OSWorld sits at 72.4%, so the gap is closing but still meaningful.

Task length and reliable delegation #

Duration is the other axis that matters, because multi-step workflows only become dependable when they can run long enough to finish the boring parts. METR has been quantifying this with a "50% time horizon" metric: the human task duration where an AI agent completes tasks with 50% success. Their March 2025 analysis shows this horizon rising exponentially with a doubling time around ~7 months.3METR, "Measuring AI Ability to Complete Long Tasks" (time horizon, ~7-month doubling)https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ If the trend from the past six years continues for two to four more years, they extrapolate, generalist autonomous agents will be capable of performing a wide range of week-long tasks. Their follow-on work explores how the horizon varies by domain and benchmark design.4METR, "How Does Time Horizon Vary Across Domains?"https://metr.org/blog/2025-07-14-how-does-time-horizon-vary-across-domains/

From a 2026 lens, Google, Anthropic, and OpenAI are all optimizing for "agents that code and operate tools for longer." Google's Gemini 3 announcement calls out 54.2% on Terminal-Bench 2.0 and 76.2% on SWE-bench Verified.5Google, "Gemini 3" announcement (Terminal-Bench 2.0 and SWE-bench Verified figures)https://blog.google/products-and-platforms/products/gemini/gemini-3/ The public Terminal-Bench leaderboard shows Gemini 3 Pro at 61.1% ± 2.8 under a specific harness and run setup, with top OpenAI and Anthropic agent runs clustered close behind.6Terminal-Bench, terminal-bench@2.0 leaderboardhttps://www.tbench.ai/leaderboard/terminal-bench/2.0

Anthropic's Claude Opus 4.5 announcement emphasizes efficiency for long runs, including 76% fewer output tokens at medium effort while matching Sonnet's best SWE-bench Verified score. At high effort, Opus 4.5 hit 80.9% on SWE-bench Verified, the first model to break the 80% threshold, and 66.3% on OSWorld.7Anthropic, "Introducing Claude Opus 4.5" (80.9% SWE-bench, 76% fewer tokens at medium effort)https://www.anthropic.com/news/claude-opus-4-5 OpenAI's GPT-5.1-Codex-Max post highlights similar dynamics: better SWE-bench Verified performance at the same "medium" reasoning effort while using 30% fewer thinking tokens, plus an "xhigh" setting for longer deliberation.8OpenAI, "Building more with GPT-5.1-Codex-Max" (30% fewer thinking tokens; xhigh effort)https://openai.com/index/gpt-5-1-codex-max/

Artificial Analysis helps keep the narrative grounded because they track the ecosystem across labs and quarters. Their reporting highlights a continued shift toward models trained and optimized for tool use and agentic execution, including reinforcement-learning-driven gains and recurring "agent stack" patterns such as sandboxes and tool registries.9Artificial Analysis, Q3 2025 State of AI Highlights Report (agentic and tool-use trend reporting)https://artificialanalysis.ai/downloads/state-of-ai/2025/Q3-2025-Artificial-Analysis-State-of-AI-Highlights-Report.pdf

Capability is not adoption #

Adoption is still noisy. Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls.10Gartner, press release on agentic AI project cancellations by 2027https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027 They also estimate only about 130 of the thousands of agentic AI vendors offer genuine capabilities; the rest is "agent washing," rebranding chatbots and RPA tools without substantive agentic features. On the flip side, Gartner predicts 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024.

Camunda's 2026 survey paints a similar picture: 71% of organizations say they're using AI agents, yet only 11% of agentic use cases reached production in the past year. 73% admit a significant gap between their agentic AI vision and today's reality. The blockers are familiar: 84% cite business risk when IT lacks appropriate controls, 80% point to lack of transparency around how AI is used, and 66% cite compliance concerns.11Camunda, "2026 State of Agentic Orchestration and Automation" reporthttps://camunda.com/blog/2026/01/closing-agentic-ai-vision-reality-gap-camunda-2026-state-of-agentic-orchestration-automation-report/