Search Is the Foundation of the Agentic Economy

Search is no longer about finding links--it’s about building the nervous system for an entire agentic economy. Will Bryk, CEO of Exa, reveals how rethinking search from the ground up for AI agents, not humans, unlocks a cascade of downstream advantages: more accurate coding agents, efficient token usage, deeper business intelligence, and even societal-level fixes like reducing polarization and loneliness. This isn’t incremental improvement--it’s a foundational shift. The implication? Companies that treat search as a commodity will be blindsided by those who see it as the core infrastructure of future workflows. This post is for builders, product leaders, and investors who want to understand where the next wave of competitive moats will form--not in models, but in retrieval.

Why the Obvious Fix--Better Models--Isn’t Enough

Everyone’s chasing bigger, smarter models. But Will Bryk sees a different bottleneck: search is the real intelligence multiplier. When Sarah Wang asks whether large language models are becoming commoditized, Bryk flips the script: “I would argue that a lot of knowledge work is actually a search problem, not an intelligence problem.” That single line reframes the entire stack.

Most teams assume that better models solve everything. They throw compute at the problem, maxing out tokens to retrieve context, summarize documents, or answer queries. But that approach hits a wall--literally. The “token apocalypse” isn’t just a cost issue; it’s a systemic inefficiency. When every input token burns budget, and every model call scales linearly with usage, the economics break.

Bryk’s insight? Use smaller, cheaper models powered by superior retrieval. Instead of embedding all world knowledge into model weights--a waste of capacity--offload that knowledge to a search system that can pull exactly what’s needed, when it’s needed. This is the “Einstein who’s never seen the world” metaphor: hyper-intelligent processing, paired with external, up-to-date tools.

"Retrieval can help solve the token apocalypse... smaller models using retrieval is much more efficient."

-- Will Bryk

This isn’t just theory. Exa’s customers see 20x cost savings by switching from brute-force model-based retrieval to Exa’s optimized search. That’s not a marginal gain--it’s the difference between a viable product and one that collapses under its own operational cost.

And here’s the kicker: this efficiency creates a second-order advantage. Because retrieval is cheaper and faster, agents can run more iterations, perform deeper validation, and maintain higher accuracy--especially in high-stakes domains like coding or investment research.

How Agents Break Human-Centric Design--and Why That Matters

Google was built for tired humans typing half-formed queries into a search bar. It’s optimized for clicks, for speed, for “good enough” answers that keep users engaged. But agents? They’re not tired. They don’t care about UI. They want comprehensive, controllable, low-latency access to all relevant information.

Bryk nails the distinction: “The world of agents searching is just completely different from human searching.” Agents have infinite patience. They don’t need summaries--they want the full dataset. They don’t simplify queries to fit a keyword box; they expand them, layering filters, domains, and semantic depth.

"An agent doesn't just want 10 pieces of information--it wants everything."

-- Will Bryk

This changes everything about system design. Human search hides controls--autocomplete, ranking, personalization--under the hood. Agent search needs those controls exposed. You want to filter by domain, enforce freshness, or run a hybrid semantic-keyword query? The agent will specify it--precisely.

And because agents are making high-stakes decisions--“finding every competitor,” “hiring the right engineer,” “debugging production code”--comprehensiveness isn’t a feature, it’s a requirement. A single missed result could mean a missed investment, a security flaw, or a failed recruitment.

This is where conventional wisdom fails. Most teams assume search is “solved” because Google exists. But Google isn’t built for this. It’s not designed to return 10,000 results with structured metadata. It’s not optimized for API throughput at planetary scale. And crucially, it doesn’t need to be--because its users are humans, not agents.

Exa, by contrast, was built for agents from day one. That means architectural choices that seem excessive for human use--extreme scalability, fine-grained filtering, high recall--are not just justified--they’re essential.

The Hidden Cost of Click Data--and Why It’s Irrelevant Now

Google’s moat has long been its mountain of click data. Every query, every click, every dwell time--fed back into ranking models to refine results. But Bryk drops a bombshell: “Human click data... just doesn’t matter that much for serving agents.”

This is a seismic shift. The asset that defined search for two decades suddenly becomes a legacy burden, not a competitive advantage. Agents don’t “click.” They don’t “dwell.” They don’t have attention spans. Their feedback loop is binary: did the result enable the correct output?

That means Exa doesn’t need to reverse-engineer human behavior. It can optimize purely for retrieval accuracy, completeness, and speed--the metrics that actually matter for agent performance.

And because of this, Exa can do with a team under 100 what once required thousands. LLMs have automated what used to be manual ranking work. A single engineer can now build a re-ranker that used to take a hundred-person team. The leverage is exponential.

But here’s where it gets harder: the quality bar is astronomically higher. Human search can tolerate 99.9% reliability. Agent search needs 99.9999999%. Why? Because agents run at scale, autonomously. A 0.1% failure rate isn’t a typo--it’s a systemic flaw that compounds across millions of queries.

Bryk puts it bluntly: “Billion-dollar investments are on the line... this has to be perfect.” That pressure isn’t a bug--it’s a feature. It forces Exa to chase “perfect search,” not just “good enough.”

Where the System Responds: Coding Agents and the Feedback Loop

One of the clearest validation points for Exa’s approach is in coding agents. These are not casual users--they’re high-precision tools where errors have real costs. And historically, their retrieval has been stuck in the “dark ages.”

But when coding agents use Exa, something changes: they make fewer mistakes, write more accurate code, and require less hallucination-checking. Why? Because Exa delivers fresher, more comprehensive, and better-structured documentation, SDKs, and technical content.

And here’s the feedback loop: better retrieval → better agent output → more trust in agents → more agent usage → more search volume → more data to improve retrieval.

This is the moat-building cycle. It’s not just about having better tech today--it’s about creating a system where success fuels further improvement. Every customer that tests Exa via A/B tests and sees better results becomes a node in this network, reinforcing the quality signal.

Moreover, this plays directly into the token efficiency story. Coding agents using Exa can use smaller models because the heavy lifting of context retrieval is offloaded. That means faster iteration, lower cost, and higher scalability--advantages that compound over time.

The 18-Month Payoff Nobody Wants to Wait For

Building a search engine from scratch is a long game. Bryk admits they started with a thought experiment: “For every query, I would take all the trillion documents on the web and run GPT-3 over it... that would be better than Google.” The problem? It would cost $10 billion per query.

So the real work wasn’t in the idea--it was in the decade-long optimization to make that vision feasible. That’s the delayed payoff: years of grinding on indexing, retrieval models, data partnerships, and efficiency--before the market even knew it needed this.

Most startups can’t wait. They chase quick wins, short-term metrics, and investor milestones. But Exa’s bet is that the winners in the agentic economy won’t be the fastest, but the most patient.

And the timeline is tightening. Bryk predicts that by the end of 2026, the shift to smaller models + retrieval will be “very noticeable.” The token apocalypse will force the industry to confront inefficiency. And when it does, the companies with mature, high-quality search infrastructure will have a decisive edge.

Key Action Items

• Reframe search as infrastructure, not a feature. Over the next quarter, audit your product’s reliance on retrieval. Is it a bolt-on, or a core system? Treat it like database or compute--because it will be.

• Optimize for token efficiency now. This pays off in 12--18 months. Start experimenting with smaller models + external retrieval. The cost savings and accuracy gains will compound as agent usage scales.

• Build for comprehensiveness, not convenience. If your agents are making business-critical decisions, demand 100% recall on key queries. This requires rethinking your search stack--start the work now, even if it feels excessive.

• Design APIs for control, not simplicity. Give agents access to filters, ranking toggles, and hybrid query modes. This creates a better user (agent) experience and reduces the need for retries.

• Invest in private, proprietary data pipelines. The next frontier isn’t just web search--it’s accessing data not on the public web. Begin partnerships or scraping initiatives to build exclusive retrieval sources.

• Run A/B tests against ground truth, not benchmarks. Industry evals are “bench maxed” and misleading. Test retrieval quality against your own high-stakes use cases--especially in coding, recruiting, and competitive intelligence.

• Hire for passion, not just skill. As Bryk says, “the world goes to who is most passionate.” In a world where tools fill gaps, motivation becomes the differentiator. Prioritize fire-in-the-eyes candidates, especially in research and engineering.