Agents in the World: Coding, Computers and Browsers
Agents moved out of demos and into daily work first in software engineering, then on screens and in browsers. This module surveys the landscape as of October 2026, explains why some domains took off faster than others, and gives practical advice for working alongside these agents.
Coding agents: the first big success
If you want to see where agents genuinely changed how people work, look at software development. In 2024, AI coding help mostly meant autocomplete. By 2026, developers routinely hand agents whole tasks (fix this bug, add this feature, migrate this module) and review the result. Several agents may run in parallel on different branches.
| Agent | From | Shape | Notes |
|---|---|---|---|
| Claude Code | Anthropic | Terminal, IDE and web | Preview February 2025, generally available May 2025, web version October 2025. Supports subagents, experimental agent teams, skills, plugins and hooks. |
| Codex | OpenAI | Desktop app, CLI and cloud | Codex desktop app in February 2026 (Windows in March), with models like GPT-5.3-Codex. OpenAI has announced plans to fold ChatGPT desktop, Codex and Atlas into one "superapp." |
| Cursor | Anysphere | AI-first editor | Cursor 3 reportedly added an Agents Window for running parallel local, worktree, remote and cloud agents (per secondary sources). |
| Antigravity and Jules | Agent-first IDE; async GitHub agent | Antigravity (November 2025) has a Manager View for parallel agents. Consumer Gemini CLI users moved to Antigravity CLI in June 2026. Jules works asynchronously on GitHub repos. | |
| Copilot coding agent | GitHub | Assign an issue, get a pull request | Works in the background inside GitHub's own workflow. |
| Open source and others | Various | CLI, IDE and frameworks | OpenHands and goose are open source. AWS Kiro, JetBrains Junie and Amp are other notable agents. Cognition's Devin and Windsurf are also in this space (reports of a Windsurf rebrand are unconfirmed). |
One harness, many surfaces
A trend worth noticing: the same agent harness increasingly runs everywhere: terminal, IDE, web, chat and background cloud jobs. Anthropic's Claude Agent SDK, for example, exposes the harness behind Claude Code so developers can build their own agents on it.
Why coding agents work so well
Coding wasn't the first domain to succeed by accident. Software gives an agent something most domains don't: a verifiable environment. The agent can act, check whether the action worked, and try again, all without a human in the loop for every step.
Fast, honest feedback
Compilers, linters and type checkers respond in seconds and don't flatter. An agent that reads the error and fixes it is doing the agent loop from Module 2 with excellent signals.
Tests as a goal
A failing test is a precise, machine-checkable definition of done. Good test suites turn a vague request into an objective the agent can verify on its own.
Git as an undo button
Version control makes every change inspectable and reversible. Branches and worktrees let several agents work in parallel without trampling each other.
Text all the way down
Code, logs, docs and config are text, which is what models handle best. Tools like grep and glob let an agent load exactly the files it needs, just in time.
This explains a lot about where agents work well elsewhere. Domains that offer quick, trustworthy checks (spreadsheets with formulas, data pipelines with validations, documents with clear rubrics) are next in line. Domains where success is subjective, slow to observe, or irreversible are much harder.
Verification can be gamed
An agent optimizing for "tests pass" can learn to weaken the tests instead of fixing the code. METR has reported examples where models attempted to cheat on tasks. Review what the agent changed, not just whether the checks went green.
Computer-use and browser agents
Not everything has an API. Computer-use agents operate software the way a person does: they look at screenshots, decide where to click or what to type, and observe the result. Browser agents do the same inside a web browser, often with extra access to the page structure.
- 1
October 2024: Anthropic computer use
Claude gained the ability to control a desktop through screenshots, mouse and keyboard actions, released as a beta for developers.
- 2
January 2025: OpenAI Operator
OpenAI's Computer-Using Agent drove a browser for tasks like booking and shopping. In July 2025 it was folded into ChatGPT agent.
- 3
2025: Project Mariner, Comet and Claude in Chrome
Google's Project Mariner, Perplexity's Comet browser, and Claude in Chrome (a pilot from August 2025) brought agents into everyday browsing.
- 4
2026: browsers with built-in agents
Gemini in Chrome added "auto browse" for paid tiers in January 2026. Claude in Chrome became generally available in August 2026, according to Wikipedia.
Computer use is slower and more fragile than calling an API: a pop-up, a layout change or a CAPTCHA can derail it. The rule of thumb is to use an API or MCP server when one exists, and fall back to computer use for the long tail of software that has no other interface.
Browsers are the riskiest environment
A browser agent reads untrusted web pages while often logged into your email, bank and work accounts. That's the "lethal trifecta" from Module 11: private data, untrusted content and a way to send data out. Use a separate browser profile, limit which sites it can act on, and confirm sensitive actions yourself.
Consolidation: the end of Atlas
Not every agent product survives. OpenAI launched ChatGPT Atlas, its own AI browser, in October 2025. On August 9, 2026 it was discontinued, and its agentic browsing features moved into the ChatGPT desktop superapp.
The lesson isn't that browser agents failed. It's that "agent" is becoming a capability of general apps rather than a separate product category. The same pattern appears in coding, where IDE, chat and cloud agents are converging into single harnesses.
Don't build on a single product
If your workflow depends on one specific agent app, expect it to be renamed, merged or retired. Keep your instructions, skills and tool connections in portable formats (AGENTS.md, Agent Skills, MCP servers) so you can switch.
Agents with their own computers
Running an agent on your own laptop has limits: it stops when you close the lid, and it shares your files and accounts. The newest pattern gives agents their own machine in the cloud, with a sandboxed environment, persistent state, and the ability to keep working while you're away.
OpenAI "dots"
Announced at DevDay on September 29, 2026: always-on agents, running on GPT-6 Astra, each with its own cloud computer. DevDay also added computer use and multi-agent tools to the Agents API.
Claude Managed Agents
Launched in April 2026 as a hosted agent harness and infrastructure, handling state, memory, permissions and scheduling. Anthropic suggests prototyping on the Agent SDK, then running on Managed Agents.
Amazon Bedrock AgentCore
AWS's agent runtime, where agents can run for up to 14 days, plus AgentCore Browser for managed browser sessions.
Giving an agent its own computer is also a security decision. A sandbox that holds only the files and credentials the task needs is much safer than an agent on your personal machine with access to everything. The trade-off is setup: you have to decide what goes into that sandbox.
Beyond code: knowledge-work agents
The harness that made coding agents work (a loop, file access, tools, skills and subagents) turns out to be general. In January 2026 Anthropic launched Claude Cowork, which brings the Claude Code style of agent to non-coding knowledge work: organizing files, assembling reports from many documents, building spreadsheets, and running multi-step research.
Knowledge-work agents face the challenge described earlier: fewer automatic checks. Nothing compiles a market analysis. Good knowledge-work setups compensate by keeping the human in the review loop, using structured outputs, and leaning on sources the agent can cite.
The same risks follow
In January 2026, security researchers reported indirect prompt-injection issues in several knowledge-work agents, reportedly including Claude Cowork, following the familiar lethal-trifecta pattern (these reports come from secondary sources). An agent that reads your documents and can send messages needs the same care as a browser agent.
Reading benchmark claims
Every launch comes with benchmark numbers. They're useful, but only if you read them carefully. Three habits help: note who reported the number, note which version of the benchmark it is, and note when.
| Benchmark | What it measures | Snapshot (October 2026) |
|---|---|---|
| OSWorld family | Completing real tasks on a desktop computer through the GUI | Human baseline on the original OSWorld: about 72.36%. Anthropic reports Claude Opus 5.5 at 81.8% on OSWorld 2.1 (vendor-reported, September 2026). OpenAI reportedly gives GPT-6 Astra 72.6% on OSWorld 2.0 (secondary). |
| Terminal-Bench 4.0 | Hard command-line tasks, with an 8-hour agent timeout | Anthropic's Opus 5.5 page lists Opus 5.5 at 66.4%, GPT-6 Astra 57.9%, Fable 5.1 55.8% and Opus 5 52.3% (vendor-reported). |
| SWE-bench Verified | Fixing 500 human-validated GitHub issues | Effectively saturated, with top scores reported in the mid-90s. Labs have moved their headline numbers to harder benchmarks. |
| SWE-bench Pro | Harder, more realistic software engineering tasks | Scores depend heavily on the split and harness, and public leaderboards conflict. Don't trust a single "#1" claim. |
Notice the comparison trap in the first row. It's tempting to say "agents beat humans at computer use, 81.8% to 72.36%," but the human baseline was measured on the original OSWorld while the 81.8% is on OSWorld 2.1, a different version. Benchmarks are versioned: when models saturate one, maintainers fix broken tasks and add harder ones (Terminal-Bench went from 2.0 to 3.0 to 4.0), so scores across versions don't line up.
Your own eval beats any leaderboard
Public benchmarks tell you roughly which models are in the frontier tier. Whether an agent works for your codebase or workflow is something only your own tasks can tell you. Module 10 covers how to build that eval.
Practical advice for working with these agents
- 1
Give the agent a way to check its work
Tests, linters, a reproduction script, an example output. The more verifiable the goal, the better the agent performs, and the less you need to watch it.
- 2
Write down the context once
Put project conventions, commands and gotchas in an AGENTS.md or CLAUDE.md file, and package repeatable procedures as skills, so you aren't re-explaining them every session.
- 3
Scope tasks like you would for a new colleague
One clear objective, the relevant files or links, constraints, and what done looks like. Large vague tasks produce large vague results.
- 4
Work on branches and review diffs
Let agents work in branches or worktrees, then review the change as you would a pull request. Check that tests weren't weakened to pass.
- 5
Limit permissions and credentials
Run agents in sandboxes or separate profiles. Don't leave production tokens where an agent can find them; real incidents have involved agents discovering unrelated credentials and deleting production data.
- 6
Run in parallel, merge carefully
Parallel agents are great for independent tasks. For tightly coupled changes, one agent at a time is usually faster overall.
- 7
Stay portable
Products merge and retire. Prefer open formats (MCP, Agent Skills, AGENTS.md) so your setup survives the next rebrand.
Key takeaways
- Coding agents succeeded first because software offers a verifiable environment: fast feedback, tests as goals, and git for review and undo.
- Computer-use and browser agents handle software without APIs, but they're slower, more fragile, and exposed to untrusted content; prefer APIs and MCP when available.
- Agent features are consolidating into general apps: ChatGPT Atlas was discontinued on August 9, 2026 and folded into the ChatGPT desktop app.
- Agents increasingly get their own sandboxed cloud computers (OpenAI dots, Claude Managed Agents, AgentCore), which helps both autonomy and security.
- Read benchmarks by source, version and date: Opus 5.5's 81.8% on OSWorld 2.1 is vendor-reported and not directly comparable to the 72.36% human baseline on the original OSWorld.
- Work with agents by giving them checks, written context, well-scoped tasks, minimal credentials and a review step.
Sources & further reading
- Claude Code changelog
- OpenAI: Introducing the Codex app
- Google Antigravity (Wikipedia)
- TNW: OpenAI shuts down the ChatGPT Atlas browser
- Decrypt: everything announced at OpenAI DevDay 2026
- Claude Managed Agents
- Amazon Bedrock AgentCore release notes
- Anthropic: Claude Opus 5.5 — Vendor-reported Terminal-Bench 4.0 and OSWorld 2.1 scores
- Terminal-Bench 4.0
- Claude (Wikipedia) — Claude Code and Claude in Chrome timeline