Operationalizing Operator - What’s still missing for the autonomous web
OpenAI’s Operator is the latest in a growing wave of AI web agents designed to automate tasks traditionally performed by humans. The rationale is clear: if machines can handle these tasks, companies can significantly cut costs or scale operations.
Now that the hype has calmed down, let’s examine how Operators fit into the broader automation landscape and why building real Agentic automation is still so difficult.
Agentic Automation Economics 101
Automation is fundamentally an economic decision - The goal is to either generate more revenue, or reduce costs. At the risk of oversimplifying, here are the few key concepts needed to understand the economics of automation:
Total Cost of Ownership (TCO)
When automating a process, the true cost extends beyond just development. It includes multiple factors:
Total Cost of Ownership = Implementation cost + (Runtime cost × Repetitions) + Maintenance cost
Any ROI assessment of automation should compare against this total cost figure.
Agentic automation unit economics
For automation to be worthwhile, it must be more cost-effective than the existing manual process. This is easily checked by plugging in the TCO values for both manual and Agentic task completion, and comparing between the two.
Operative envelope
AI agents operate within specific execution constraints. While browsers offer a flexible way to perform tasks, agents like Operator still face limitations, including:
CAPTCHA and 2FA challenges
Identity and access management
Restrictions on running inside customer cloud environments
Data processing constraints
How Operator affects the automation unit economics
Properly using Operator can reduce some of the Implementation and Maintenance costs, but it doesn’t eliminate them. As automation technology evolves, the goal remains to minimize TCO and make automation viable for a broader range of tasks.
What’s left? The Gaps in Operationalizing AI Agents
For production-grade automation, several critical components are still missing. These gaps fall into three main categories:
Task planning gaps
Inputs Structure Definition: Figure out the data or parameters your automation requires.
Operator simplifies some aspects of input definition by using natural language interfaces, but structured inputs are still necessary for robust automation.
Outputs Structure Definition: Establish the expected results or deliverables.
While Operator provides standard output formats, ensuring the correct output structure for business applications remains an additional step.
Set Success Criteria: Know how you’ll measure whether automation is working.
Operator introduces basic validation mechanisms but lacks deep success criteria validation measures.
Security & Disaster Planning: Understand how to protect against failures or malicious attacks.
Web agents including Operator are hard to guardrail properly while still giving them enough access to complete their tasks.
Implementation gaps
Transition from Code to “Self-Healing”: Build solutions that adapt to changes without breaking.
Current web agents including Operator do not transition to deterministic code based execution
Task Memorization: Reuse a single successful task completion as a base for future executions of the same flow.
Current web agents including Operator do not “learn” how to perform a specific task
Maintenance gaps
Observability & Monitoring: Keep tabs on performance and any anomalies. For hundreds of defined tasks and millions of task executions.
Operator provides minimal observability beyond task execution logs, making it hard to debug failures at scale.
Self-Improving: Over time, defined tasks should:
Eliminate unnecessary steps.
Shift from browser-driven to API-driven workflows where possible.
Cache data for faster performance.
Current web agents including Operator do not self-improve
Closing thoughts
Although Agentic automation is improving rapidly, there is still much more left to do.
Anchor Browser is the purpose-built infrastructure layer for Agentic automation, solving issues around - Captcha solving, Identity management, Monitoring, Deterministic execution, and more.
If you are:
Building web-based AI agents
Solving one of the operational gaps mentioned above
A service provider looking to enhance web automation
Seeking to minimize TCO while scaling automation
Schedule a call with Anchor founders to see how we can help.