EU AI Act high risk obligations are now enforceable. Check your exposure
Insights About us Careers
Contact us
AI Agents & Agentic Automation

Browser and Computer-Use Automation Agents

Agents that operate a browser or desktop application when there is genuinely no API to call, engineered with isolated credentials, defence against instructions embedded in web pages, and a clear-eyed view of how brittle this is.

5 to 9 weeks
Typical build
Fixed scope
Commercial model
Last resort
By design

We treat browser automation as the option of last resort, and we will spend the first week trying to talk you out of it. Interfaces change without notice, credentials have to live somewhere, and a web page the agent reads can contain instructions aimed at the agent. When there is genuinely no alternative, it is buildable, and it needs to be built carefully.

In one paragraph

Browser and computer-use automation agents operate software through its user interface rather than its API: navigating pages, reading the screen, filling forms and clicking, driven by a model that decides what to do next, used where a system exposes no programmatic interface and cannot be changed.

Try these before automating a browser

  1. An API you have not found. Many systems have one that is undocumented, unadvertised or available on request from the vendor.
  2. The network calls the interface itself makes. Frequently a supported, stable path that avoids driving the interface at all.
  3. Bulk export and import. Unglamorous, robust, and sufficient for a surprising share of what gets proposed as automation.
  4. A database or reporting connection. For reads, often available where an API is not.
  5. Asking the vendor. Integration is frequently on a roadmap that a customer request will accelerate.

If none of those work, browser automation is legitimate, and the rest of this page is about doing it properly.

How we build browser agents safely

Isolate credentials and the environment

The agent runs in a contained environment with dedicated accounts, least-privilege permissions, and credentials injected at runtime from a secret store rather than stored in a profile. It never runs on a person's machine under their identity, which is how these projects create audit problems that outlive the automation.

Treat every page as hostile input

Text on a page can contain instructions aimed at the model. We separate page content from instructions, restrict what page-derived content is permitted to influence, keep the tool surface minimal, and test with deliberately poisoned pages. An agent with browser access and broad tools is an unusually attractive target.

Prefer deterministic steps

Where a step can be a stable selector or a recorded interaction, it is, and the model is used only where the interface genuinely varies. A fully model-driven interaction is slower, more expensive and more likely to do something unexpected than a scripted step with a model fallback.

Verify the outcome, not the click

After every consequential action the agent confirms the resulting state: the record shows the new value, the confirmation number exists, the total matches. Clicking a button is not evidence that anything happened, and this check is what turns a fragile script into something you can leave running.

Expect breakage and detect it fast

Interfaces change. We build canary runs that exercise the critical path on a schedule, alert on failure with a screenshot and the step that broke, and pause the automation rather than continuing blindly. Ongoing maintenance is a real cost and we say so before you commit.

Worth knowing

Check the terms before you build

Automated access is prohibited or restricted by many services' terms, and for third-party systems this is a contractual and sometimes legal question rather than a technical one. We raise it in week one, in writing, because discovering it after launch is considerably more expensive.

What this is good for, and what it is not

SituationSuitabilityWhy
Internal legacy system, no API, stable interfaceGoodChange is controlled and breakage is predictable
Supplier portal you must use dailyWorkableContractually permitted, though brittle and needs canaries
Public site with anti-automation measuresPoorTerms, blocking and an arms race you will not win
High-volume data extractionPoorSlow and expensive per record; find another route
Occasional multi-step task across two systemsGoodLow volume, high manual cost, tolerable fragility
Process

How the engagement runs

The first week is spent looking for a route that does not require driving an interface at all.

Week 1

Alternatives and terms

We look for an API, network path, export route or vendor option, and review the terms of every system in scope. Findings are written down before anything is built.

Weeks 2 to 3

Environment and credentials

Contained runtime, dedicated least-privilege accounts, secret management and full session recording.

Weeks 4 to 6

Task automation

Deterministic steps where possible with model fallback, outcome verification after every consequential action.

Weeks 7 to 8

Hardening

Injection testing with poisoned pages, failure handling, canary runs and alerting with screenshots.

Week 9

Handover

Runbooks, the maintenance schedule and an honest estimate of ongoing upkeep as interfaces change.

Deliverables

What you receive

A working automation, contained, monitored, and delivered with an honest maintenance forecast.

01

Alternatives assessment

What else was tried and why it did not work, plus the terms review for every system in scope.

02

Contained runtime

Isolated environment, dedicated accounts, least-privilege access and runtime secret injection.

03

Automation

Deterministic steps with model fallback, outcome verification and full session recording.

04

Injection defence report

Poisoned-page testing, the separations applied, and the residual risk stated plainly.

05

Canary monitoring

Scheduled critical-path runs with screenshot-attached alerting and automatic pause on failure.

06

Maintenance forecast

Expected breakage rate, upkeep effort and the trigger for revisiting the approach.

Fit check

Is this the right engagement?

Worth being direct. Browser and Computer-Use Automation Agents is the wrong spend in some situations, and those are listed rather than buried.

Good fit if

  • The target system genuinely has no API and cannot be changed.
  • Automated access is permitted by the terms, and you have confirmed it.
  • The task is multi-step, repetitive and currently costs real time.
  • You accept ongoing maintenance as interfaces change.
  • Dedicated service accounts can be created for the automation.

Choose something else if

  • An API exists, even an undocumented one. Use it.
  • The terms of service prohibit automated access.
  • The volume is high enough that per-record cost and fragility become the story.
  • Nobody will maintain it when a supplier redesigns their portal.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

When should we use a browser automation agent?

Only when there is no API, no supported network path, no export route and no vendor option, and the terms permit automated access. It is a legitimate last resort for internal legacy systems and mandatory supplier portals, and a poor choice almost everywhere else.

How fragile is browser automation really?

Fragile enough that monitoring is part of the build rather than an option. Interfaces change without notice, so we use deterministic steps where possible, verify outcomes rather than clicks, run canaries on a schedule and pause automatically on failure. Expect real maintenance and budget for it.

Can a web page attack the agent?

Yes. Text on a page can carry instructions aimed at the model, and an agent with browser access and broad tools is an attractive target. We separate page content from instructions, keep the tool surface minimal, test with deliberately poisoned pages and document the residual risk.

Where do the credentials live?

In a secret store, injected at runtime into a contained environment running under dedicated least-privilege accounts. The agent never runs on a person's machine under their identity, which keeps the audit trail clean and the access revocable.

Is this legal?

It depends on the system and its terms, which is a contractual question we raise in week one and put in writing. For internal systems it is usually straightforward. For third-party portals it needs checking before anything is built, and we will not proceed on an assumption.

Is this the right engagement?

Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.