AI coding agents and developer workflows (Claude Code & Agent tools)

AI coding agents matured into practical tools. - Simon Willison published a year-end long review on AI in 2025 - 2025 key breakthrough: reasoning models enabled Agents - Examples of reasoning models: OpenAI o1, DeepSeek R1 - Reasoning models use internal stepwise thinking and verification - Agents + code-execution sandboxes can write, run, debug, iterate - AI search evolved into usable research assistant (GPT-5 Thinking, Google AI Mode) - Anthropic released Claude Code CLI in Feb, embedded in Sonnet release - Claude Code CLI generated $1B annualized revenue by year-end - Major vendors released CLI agents: OpenAI Codex CLI, Google Gemini CLI - Asynchronous cloud agents (Claude Code for Web, Codex Cloud, Google Jules) produce PRs - ChatGPT image editing launched in March, gained ~100M users in a week - Google Nano Banana focused on professional image/infographic generation - Simon: Chinese open models topped open-weight rankings at year-end - Top Chinese models listed: GLM, Kimi K2, MiMo, DeepSeek V3.2, MiniMax - DeepSeek R1 release triggered ~$600 billion USD instantaneous NVIDIA market cap drop - Safety concerns: risk-normalization likened to Challenger disaster warnings - Conformance suites/tests recommended as infrastructure for adoption - Ben Tossell used AI Agents (≈3B tokens) to build multiple projects without traditional coding - Ben’s workflow: terminal CLI, context provision, spec mode, high-agent autonomy, iterative testing - Lee Robinson relies on Agents but enforces multi-stage human review and testing tools - Cursor in-browser agent control, Plan Mode, Debug Mode, and automated logging improve workflows - Recommended engineering practices: type systems, linters, tests, modern toolchains (tsgo, biome/oxlint, Bun) - Mobile-first coding noted: significant coding done on phones with Claude Opus 4.5 and async agents - Metaphor: Claude and Claude Code are a matched “sword and swordsman”; models differ in tooling fit - Some users report planning large Claude Agent projects but prefer manual implementation until scaling needed
Key insights
Reasoning models made agents practically useful
Reasoning models were identified as the 2025 technical breakpoint that enabled coding agents to perform reliably. Reports cite models such as OpenAI o1 and DeepSeek R1 and describe a trained internal stepwise thinking process: decomposing problems, attempting solutions, checking results, and iterating. Training on tasks with automatic verification (math, programming) produced learned multi-step solution strategies. When combined with execution sandboxes, these reasoning-capable models can autonomously write code, run tests, inspect errors, and iteratively fix bugs until resolved, turning agents into practical contributors rather than brittle chat tools.
Claude Code CLI achieved major commercial traction
Anthropic’s Claude Code CLI, quietly released in February within the Claude 3.7 Sonnet announcement, generated an estimated $1 billion in annualized revenue by year-end. The CLI form factor spurred competitors to release similar tooling—OpenAI Codex CLI, Google Gemini CLI, and offerings from Qwen and Mistral. Asynchronous cloud agents (Claude Code for Web, Codex Cloud, Google Jules) were reported to accept tasks and return pull requests, enabling users to dispatch jobs from phones and receive completed PRs minutes later. The CLI and async-agent model became a high-value product vector.
Non-developers can build real projects using AI Agents
A practitioner who does not identify as a traditional programmer described producing multiple deployed projects using AI Agents and approximately 3 billion tokens of model interaction. Projects included personal CLI sites, an open social tracker, a year-end summarization product, multiple CLI utilities, a crypto trading tracker, an AI-generated demo video system, Telegram bots, and other experiments. The workflow emphasized a terminal/CLI interface for observability, providing context (docs, repos), entering a spec mode for planning, assigning high autonomy to agents, and iterative testing and refinement.
Human oversight and multi-stage review remain essential
Experienced developers reported delegating most code generation to AI Agents while retaining strict human control over review and acceptance. Typical practice includes real-time inspection during generation, editor-integrated AI review before PR submission, and automated final checks (e.g., Bugbot) on PRs. Recommended engineering safeguards include explicit, testable goals, type systems, linters, and robust test suites. Developers emphasized planning (Plan Mode), systematic debugging (Debug Mode with logs), and staged review to avoid regressions and ensure production-quality output.
In-browser agent control and workflow modes speed development
Integrated browser-capable agents improve debugging and feature development by directly accessing network requests, console logs, and page elements. Tools with Plan Mode and Debug Mode facilitate structured planning and automated log insertion to reproduce issues; agents can analyze logs and iterate fixes. Practitioners recommended starting projects simply and incrementally adding structure as complexity grows. Tooling choices cited include Cursor’s in-browser agent, and modern toolchains such as tsgo, biome/oxlint, and Bun to improve developer productivity.
Ecosystem shift: open Chinese models and safety concerns
By year-end, the top open-weight model rankings reportedly listed five Chinese models—GLM, Kimi K2, MiMo, DeepSeek V3.2, and MiniMax—and DeepSeek R1’s release coincided with an estimated $600 billion USD single-day loss in NVIDIA market capitalization. Safety researchers warned of a 'risk-normalization' reminiscent of pre-Challenger indications, noting common patterns of unmonitored agent autonomy ('YOLO' mode). Conformance suites and language-agnostic test sets were recommended as foundational infrastructure to validate agents and new protocols.
Mobile and CLI-first workflows became mainstream for agents
Practitioners reported substantial coding activity conducted on phones using Claude Opus 4.5 and asynchronous cloud agents, including completing a C-to-Python library port on an iPhone. The CLI was highlighted as a powerful interface for observing agent actions and orchestrating tasks. Some users expressed that Claude Code–style agents can perform many jobs and that human identity could be exploited for work application, while others planned major Claude Agent projects but chose manual implementation until system bottlenecks required scaling.