Anthropic Expands Claude Computer Use With Browser, Skills and File Storage — What It Means for Back-Office Agents

Anthropic just took Claude out of 'answer my question' mode and put it properly to work inside your applications. On 20 August 2026, computer use, the Skills API and the Files API all moved to general availability on the Claude Platform, and a new browser use tool was added that lets Claude drive a web app the way a trained employee would — reading the screen, filling forms, clicking buttons, and saving the result. For an Indian business running back-office work in Tally, a CRM, a vendor portal or an internal dashboard, this is the first version of Claude that can genuinely do the work, not just describe it. The catch is that 'can do the work' and 'should be trusted to do the work unsupervised' are two very different things, and that gap is where most companies will get into trouble first.

Three previously-beta features lost their beta headers and became production tools on the Claude Platform: computer use, the Skills API, and the Files API. Alongside them, Anthropic shipped a new browser use tool that extends computer use from controlling a full desktop to controlling a browser viewport inside a customer's own application. Anthropic has also made Skills and Files available through Microsoft Foundry, which matters for any Indian enterprise already standardised on Microsoft's AI tooling for procurement or compliance reasons. In plain terms: these are no longer experiments you bolt on with a warning label. They're meant to run in production, at scale, inside real workflows.

Computer use gives Claude the ability to read and write files, take screenshots, and control a mouse and keyboard on a machine, using a toolset Anthropic identifies as computer_toolset_20260801. The GA version supports batch actions, meaning Claude can execute several steps — click, type, screenshot, click again — in a single model turn instead of one action per round trip. That sounds small, but it's the difference between an agent that takes minutes to fill a ten-field form and one that does it in a handful of turns. Think of a back-office example: matching a vendor invoice PDF against a purchase order in an ERP screen, flagging mismatches, and updating a status field. Computer use is what makes that a single automated pass instead of a person tabbing through three windows.

Browser use, running on browser_toolset_20260801, is the feature Channel Insider calls the 'key addition' in this release. Instead of controlling the entire desktop, it operates inside a browser viewport that the application itself hosts, and it can read the page's accessibility tree, forms, elements and open tabs — not just a screenshot. That means Claude gets element references it can act on directly, can manage tabs, report downloads, and optionally accept file uploads, on top of the older screenshot-and-click approach. Practically, this is what lets an agent log into a government portal, a courier tracking page, or an internal web dashboard and complete a task that has no API — GST return status checks, tracking number lookups, or pulling a report from a SaaS tool that only exposes a UI, not an endpoint. A huge amount of Indian back-office work still lives behind exactly this kind of login-and-click interface, which is why browser use matters more here than it might in a fully API-first business.

The Skills API has been simplified specifically for uploading and versioning your own skills, built around folder-based packages of instructions, scripts and templates, using a SKILL.md model similar to Claude Code's filesystem skills, and executed in a sandbox. In effect, a skill is your team's standard operating procedure written down once and attached to every relevant request, instead of re-explained in every prompt. A finance team could define a 'monthly reconciliation' skill describing exactly which fields to check, which exceptions to escalate, and which format the output report should take, then version it as the process changes. This is a meaningful shift from prompt engineering to actual process engineering, and it's the piece that lets an agent behave consistently across hundreds of runs instead of drifting each time someone tweaks the wording.

The Files API now runs on a stable /v1/files endpoint, supports expiration controls via expires_in_seconds and expires_at, offers paginated listing, and — the headline number — gives organisations up to 1 TB of storage with rate limits five times higher than before. Anthropic built this specifically for long computer-use trajectories, which accumulate screenshots and intermediate documents fast. Upload a file once, reference it by ID afterwards, and an agent can pull the same invoice, contract or product catalogue across dozens of steps without re-uploading it each time. For any workflow that involves reading a long PDF policy document, saving a generated report, or carrying a document through a multi-step approval chain, this quietly removes what used to be a real bottleneck.

Three failure patterns show up almost every time a team wires up computer use or browser use without proper scaffolding.

The technology to give an AI hands and eyes on your systems is here. The discipline to do it safely still has to be built by hand, every time.

None of this means the capability isn't worth using. It means the sensible version of this is a bounded agent — scoped to specific screens and specific actions, logged at every step, and stopped before anything irreversible, with a human in the loop wherever money, data deletion, or customer communication is involved.

This is precisely what ODIV's multi-agent-systems service is for. Instead of switching Claude's computer use and browser use tools on and hoping for the best, ODIV's engineers design the agent's boundaries first: which screens it can enter, which actions are allowed, which fields it can write to, and where it must stop and wait for your team's approval. We then build the actual skill files for your specific processes — say, reconciling delivery challans against a courier portal, or pulling monthly GST filing status across multiple client accounts — with the Files API handling document storage and expiry, and full logging so every click and file the agent touches is reviewable afterwards. Our team works hands-on in the same generation of AI build tools this technology comes from — Lovable, Claude Code and similar AI coding environments — combined with traditional engineering discipline for the parts that need to be solid: authentication, error handling, approval workflows, and integration with your existing ERP, CRM or internal tools. That combination is what lets us get a working, properly scoped agent live in a fraction of the time and cost of a traditional custom-coded automation project, because we're not billing for every line of boilerplate by hand. If some part of the same workflow also needs to notify your team or a customer once a task completes, ODIV Engage can send that update straight on WhatsApp, but the agent build itself sits under multi-agent-systems. If you're looking at Claude's new browser and computer use tools and wondering what a safe, production version of this looks like for your business, start a chat with us on WhatsApp and we'll walk through what a bounded agent for your specific process would actually take to build.