A browser agent reads pages, clicks controls and fills forms through the same interface a person uses. It is useful when an application has no API, or when a bank, courier or supplier portal is the only route to a task. It is also exposed to every change a human user would notice: a redesigned button, a new cookie banner, an expiring session or a challenge page. Many production failures are state and control problems, not evidence that the model is too weak.
Start with a diagnostic method, not a guess
Before changing the prompt, capture four artefacts: the run trace, a screenshot and DOM snapshot at the failure point, the exact input that started the run, and the downstream system state. Then classify the failure. Is the agent seeing the wrong element, losing its session, being blocked by policy, receiving a downstream error, or reasoning incorrectly? Reproduce the run with the same account and data. Note whether it fails every time or only under load. That classification determines the fix; rewriting instructions usually does not.
Symptom 1: the agent clicks the wrong control
Likely cause: selectors built on page position or CSS classes, duplicate labels, a hidden mobile menu, or a modal that appears after the agent has already chosen a target.
Fix: locate elements by accessible role and name where possible, add stable test identifiers with the application team, and assert the page title and step before every action. Take a screenshot immediately before the click.
Prevention: treat the workflow as an integration contract. When the interface changes, run the agent against a staging copy before release and notify the automation owner.
Symptom 2: the session expires halfway through
Likely cause: the identity provider session is shorter than the task, a second factor is requested mid-run, or the agent is using a personal account with device limits.
Fix: give the agent a dedicated service identity with the minimum permissions it needs. Add a re-authentication checkpoint, save progress before sensitive steps, and route the run to a person when the login page appears.
Prevention: compare typical task duration with session lifetime, alert on unexpected login redirects, and avoid sharing a human administrator account.
Symptom 3: the same order is submitted twice
Likely cause: a timeout triggers a retry even though the first submission succeeded, a double click is interpreted as two actions, or the target system has no idempotency control.
Fix: stop on an unknown result instead of retrying blindly. Read the order list or transaction history to confirm what happened. Where an API supports an idempotency key, use it. Otherwise send the case to a human review queue.
Prevention: enforce a single writer for each case, reconcile submitted records against the system of record, and keep a kill switch that halts new runs immediately.
Symptom 4: the run says success but nothing changed
Likely cause: the agent treats a success toast as proof, or waits for a network-idle signal that never arrives and then reports completion anyway.
Fix: verify the outcome in the system of record, not on the page that accepted the action. Assert on a persisted status, reference number or updated record. Capture that evidence in the run log.
Prevention: write acceptance checks before launch, include a canary record that is checked after every run, and report unknown outcomes as failures rather than successes.
Symptom 5: a bot challenge blocks the workflow
Likely cause: the target site uses a challenge to enforce its terms, or the automation pattern looks abusive to its defences.
Fix: stop. Use an official API, a partner integration or a sanctioned bulk-export route. If none exists, keep the step manual or ask the vendor for written permission. Working around the control creates legal and operational risk.
Prevention: review the target site terms during procurement, and record which systems are approved for automation before a workflow is built.
Symptom 6: the agent loops until the budget is gone
Likely cause: retries that do not change state, ambiguous instructions, pagination that never ends, or an error page that looks like a valid step.
Fix: add a step budget, a no-progress detector and a watchdog that stops the run after a fixed time. Escalate with the trace rather than letting the loop continue.
Prevention: set per-run limits for steps, time and spend. Alert when a workflow uses most of its allowance, and review the longest runs weekly.
A triage order for the first fifteen minutes
- Confirm whether the target application is healthy for a human user.
- Compare the last successful run with the current page or API response.
- Check authentication and session state before touching the prompt.
- Read the trace for a repeated action, a skipped step or an unhandled modal.
- Query the system of record to establish what actually changed.
- Disable the workflow if money, customer data or stock is at risk, then investigate.
Controls that belong in production
- Scoped service accounts with multi-factor authentication where the target supports it.
- An allowlist of domains, actions and data fields the agent may use.
- Human approval for payments, refunds, deletions, contract changes and outbound messages.
- Immutable run logs with screenshots and inputs, redacted before storage.
- Rate limits, concurrency limits and a documented kill switch.
- Change management for the model, browser and target application.
- An incident runbook that names the person who can stop the workflow.
What to measure after each fix
Track success rate by workflow, the share of runs needing human intervention, duplicate or incorrect transactions, time to detect a failure, and cost per completed task. Add a count of runs blocked by policy, because that number reveals whether the workflow is fighting the target site. Review these measures after every interface or model change. A fix that raises the success rate but also raises duplicate submissions is not an improvement.
Before a browser agent touches a live process, define its owner, permissions and stop button. Our business automation team builds those controls into the workflow. The pricing page explains the engagement models, and the AI automation evaluation guide lists the vendor questions worth asking first.
Sources
- OWASP, Top 10 for Large Language Model Applications, accessed 17 November 2026.
- NIST, AI Risk Management Framework, accessed 17 November 2026.
- Playwright, Best Practices, accessed 17 November 2026.
- Gartner, Applying Uniform Governance Across AI Agents, accessed 17 November 2026.
Responses (0)
No comments yet. Add the first one below.
Leave a response
Your email address will not be published. Required fields are marked *.