The honest answer: today’s frontier models can hold close to a million tokens of context, chain multiple tools together without you babysitting every step, and click through a real screen like a person would, but each of those breaks in predictable ways worth knowing before you build a workflow around them.
| Context window | Roughly 1M tokens now standard on leading models |
|---|---|
| Tool chaining | Multi step, self verifying loops, not single calls |
| Computer use | Screenshot based, coordinate clicks, no direct page access |
| Biggest myth | Fully unsupervised all day agent runs |
What actually changed with long context?
A year ago, “long context” meant you could paste in a long document and hope the model remembered the middle of it. Today the leading models genuinely hold hundreds of thousands of tokens, and several now sit at roughly a million as a standard offering, not a gated beta. That’s enough to load an entire codebase, a quarter’s worth of customer transcripts, or a full book and ask real questions against all of it at once.
Here’s what’s different in practice: retrieval across that whole window is much more even now. The old “lost in the middle” problem, where models nailed facts near the start or end of a document but fumbled anything buried in the center, is measurably better. It is not gone. If you drop in 300 pages and ask a question that depends on connecting a detail on page 40 to a detail on page 260, run the test yourself before you trust the answer in anything that matters. I do this with my own documents before I let an agent act on them: plant a fact I know is true somewhere in the middle third of the file and ask a question only answerable by finding it. If it misses, I chunk the document instead of trusting the full dump.
Can these models actually use tools reliably now?
Yes, and this is the biggest quiet shift of the last year. Tool use used to mean a model could call one function if you prompted it carefully. Now it can plan a sequence: search, read the result, decide it needs a second search, call a calculator, write a file, and check its own output, all without you writing glue code for each step. At Quarry, the agents doing marketing work daily aren’t calling one API and stopping. They’re deciding which of several tools to reach for based on what the last tool returned.
The exact wording matters more than people expect. Tell a model “use the search tool if needed” and it will often skip it. Tell it “before answering, call search and cite what you find, then verify the citation matches the claim” and completion rates on multi step tasks go up noticeably. Naming the verification step explicitly, not just the action, is what turned tool use from a demo into something you can run unattended for a few steps at a time.
What is computer use, and does it actually work?
Computer use means the model looks at a screenshot, decides where to click or type, and acts on a real interface the same way you would, no API required. This was barely functional outside a curated demo a year ago. Today it can genuinely fill out a form, navigate a dashboard, or complete a purchase flow end to end.
Where it still fails: it’s reading pixels, not the underlying structure of the page, so it moves slower than a script and it can misjudge a click when a layout shifts, a modal pops up unexpectedly, or two buttons look nearly identical. I do not let it run more than a handful of actions without a checkpoint. Screenshot, confirm the state matches what was expected, then proceed. That single habit is the difference between a computer use agent that quietly does the wrong thing for ten minutes and one you can actually trust with something that matters.
Is “just let it run all day” a myth?
Yes, plainly. This is the most shareable thing in this piece because it’s the claim I hear most and believe least. Long, fully unsupervised agent runs, the kind marketed as “give it a goal and walk away,” still drift. Small errors compound: a misread screenshot, a tool call that returns a plausible but wrong result, a piece of context that fell just outside the window. None of that shows up as a crash. It shows up as a confidently wrong output three hours later that looks fine until you check it.
What actually works, and what I run across my own brands, is agents that do real, bounded work with a checkpoint at the end of each unit: research this, draft that, execute this one transaction, then hand back for a look before the next unit starts. That is not a lesser use of these models. It is the difference between using the tool for what it’s good at today and pretending it’s something it will be next year.
What should you actually try this week?
- Test long context on your own documents with a planted fact before trusting it on anything that matters.
- Rewrite your tool prompts to name the verification step, not just the action.
- Add a screenshot checkpoint every few actions in any computer use workflow.
- Treat “fully autonomous all day” claims as marketing until you’ve watched the drift yourself.
Takeaways
- Before trusting long context on any document that matters, plant a fact in the middle third and test retrieval yourself first
- Rewrite tool use prompts to explicitly name the verification step, not just the action, to raise multi step completion rates
- Add a screenshot checkpoint every few actions in a computer use workflow instead of letting it chain unsupervised
- Treat fully autonomous all day agent claims as marketing until you’ve watched the drift firsthand
- Break agent work into bounded units with a checkpoint at the end of each rather than one long unsupervised run
What happens next
Context windows will keep growing, but the real fight from here is retrieval accuracy across the whole window, not the size of the number on the spec sheet. Watch whether computer use starts reading page structure instead of just pixels, since that would fix its most common failure mode. I’ll keep testing each new release against my own documents rather than trusting the benchmark.
Questions people actually ask
How big is the context window on top AI models right now?
Leading models now offer roughly 1 million tokens of context as a standard feature rather than a limited beta, enough to hold an entire codebase or a full book. Retrieval across that space is more even than a year ago, but accuracy still drops for details buried in the middle of very long documents, so testing on your own material before trusting it still matters.
Can AI agents actually chain multiple tools together on their own?
Yes, this has shifted noticeably. Models can now plan a sequence of tool calls, search, read results, decide a second call is needed, then check their own output, without step by step glue code from a developer. How reliably this works depends heavily on prompt wording, especially whether the verification step is spelled out explicitly rather than implied.
What exactly is computer use in AI and how does it work?
Computer use lets a model look at a screenshot of a real screen, decide where to click or type, and act on an interface the same way a person would, without needing an API connection. It works from pixels rather than the page’s underlying structure, so it is slower than a script and can misclick when a layout shifts or a modal appears unexpectedly.
Can I really just give an AI agent a goal and let it run all day unsupervised?
No, this is considered a myth by practitioners who use these tools daily. Long unsupervised runs tend to drift as small errors, like a misread screenshot or a plausible but wrong tool result, quietly compound into a confidently wrong output hours later. The more reliable pattern is bounded units of work with a checkpoint after each one.
How do I stop a computer use agent from clicking the wrong thing?
Add a screenshot checkpoint every few actions rather than letting the agent chain many steps unsupervised. Confirming that the screen state matches what was expected before proceeding catches misclicks caused by shifted layouts, unexpected modals, or near identical buttons before they cause further errors downstream.
Sources
- Claude Opus 4.6Anthropic
- Computer use toolAnthropic (Claude Platform Docs)
- Lost in the Middle: How Language Models Use Long ContextsarXiv / Transactions of the Association for Computational Linguistics
- Building Effective AI AgentsAnthropic