This is Part 2 of our series on unlocking 100% of the internet. [Read Part 1 here.]
In Part 1, we established the problem: search engines index less than 5% of the internet. The remaining 90+% lives behind logins, workflows, forms, and interactive interfaces that crawlers can't navigate.
We also showed why the "agentic" solutions don't work: agentic search improves semantic retrieval of the same 5%, and browser agents automate navigation of that same 5%, but neither unlocks the 90%.
The solution isn't better indexing or faster automation. It's operational navigation, treating the web as a dynamic environment to be explored and acted within, not a static corpus to be cataloged.
Google crawls, operational web agents, well, operate!
An operational web agent doesn't wait for content to be indexed. It goes directly to the source and interacts with it:
In short, it replicates the intentionality of human browsing, at computational scale.
For example: finding every clinic in a city that accepts online bookings. A search engine would look for indexed pages mentioning "clinic" and "online booking." An operational web agent would visit local directories, detect booking widgets, traverse calendars, and extract verified availability data directly from the interface, like a human researcher operating hundreds of windows simultaneously.
The value isn't speed on individual tasks. A human researcher takes 30 seconds to check a clinic's availability. An operational agent completes the same task in 25 seconds. That's marginal. The competitive advantage lies elsewhere entirely.
Operational web agents can run hundreds of thousands of parallel sessions simultaneously, executing what would require an army of researchers working in perfect coordination.
A single human checking 100 salon websites for availability would need hours. An operational web agent completes it minutes. A service aggregation company doing the same across 40,000 websites across the world is impossible to even imagine. But, a TinyFish customer today performs that action a few times every single day.
The architecture makes this possible: visual and programmatic understanding of web interfaces, decomposition into task-specific micro-agents via fine tuned models, and browser infrastructure designed for massive scale with fingerprinting and anti-detection capabilities.
Humans make errors when processing the 47th supplier portal of the day. Attention drifts. Pattern matching degrades. Subtle discrepancies get missed.
Operational web agents maintain judgment consistency across thousands of complex workflows, catching discrepancies, applying nuanced business logic, and flagging anomalies that human fatigue would miss.
The value proposition: parallel scale for tasks humans can't sustain, and error reduction for reasoning that deteriorates with repetition.
That's what "human-like navigation at beyond human scale" means.
Most browser agents take screenshots continuously and run inference at every step, expensive, slow, and brittle. This is why all current versions of agentic browsers have ended up being curiosities rather than production systems.
Operational navigation requires a fundamentally different architecture.
When encountering a new site, the system uses multi-modal reasoning at first pass, it sees what a human sees, as well as what a machine should see. It understands layout, identifies interactive elements, and maps the navigational surface.
It's like an experienced doctor examining a patient visually, then checking the X-ray and MRI to get the full picture.
Instead of re-analyzing at every step, the learning system codifies navigation patterns into efficient execution paths. This is where specially fine tuned models operate, specialized sub-agents handle different interaction types: forms, calendars, search interfaces, authentication flows.
The system codifies not just actions, but also the validation checks needed to detect site changes or unexpected conditions. This dual codification, execution plus verification, makes sub-components reusable across multiple projects rather than single-use scripts. When a check fails, the system falls back to multi-modal reasoning to adapt.
These codified patterns execute locally through EdgeLearn, a deterministic web execution and reasoning engine built on OpenEvolve. EdgeLearn runs workflows at machine speed while continuously validating for inconsistencies.
When it encounters unexpected patterns or site changes, it escalates to the multi-modal browser learning agent to adapt and update its codified knowledge. This local-first, learn-when-needed architecture handles navigational complexity an order of magnitude beyond traditional RPA.
When sites redesign interfaces, the system doesn't break silently. It recognizes uncertainty, falls back to multi-modal reasoning, re-learns the new pattern, and updates its codified paths.
Recovery happens in minutes, not months. It's not a perfect adaptation, it's engineered resilience at scale.
Beyond immediate operational intelligence, the architecture produces valuable data assets. One work product of operational navigation is flattened websites.
For example, serving one enterprise customer, we've collected over a billion data points of hotel prices across Japan. This is data from pages that will never be seen again because they're ephemeral. This data is valuable for business decision-making, but also for making models and agents perform and reason better through reinforcement learning.
The stack, visual understanding, learned codification, local execution, massive parallelization, and RL based on flattened web data, makes operational navigation viable at enterprise scale.
Most enterprises don't need to map the entire internet. They need to understand and act within their operational domains: supply networks, compliance portals, partner ecosystems.
That's where operational web navigation fits: bounded operational intelligence, because context constrains complexity.
Google needs global scale because it serves everyone. Operational navigation doesn't.
Each enterprise use case, healthcare, logistics, compliance, real estate, is self-bounded. The agent doesn't need to "understand the web," only the logic of its operational surface.
Within those constraints, operational systems outperform search in precision, trust, and usefulness. Operational web navigation within domains gives enterprises something search never could: verified, explainable intelligence about systems they already use but can't fully see.
The architecture exists. The technology works. Operational web agents can navigate authenticated systems, execute multi-step workflows, and extract structured intelligence at scale.
But what does this actually look like when deployed across industries? How do enterprises use operational navigation in practice? What use cases unlock first? What makes this defensible as a business model?
[Part 3 will show operational navigation in action, from e-commerce to healthcare to competitive intelligence, and explain why this is hard to replicate.]
