Agent Discovery: New Game, New Rules

The case for Agent Discovery Optimization (ADO)

Back to the Future

Imagine it’s 1998, and you are seeing this new thing called Google (or is it Googol) emerge; everyone’s talking about how it is miles ahead of Yahoo, AltaVista, Lycos, etc. Now, imagine it’s 2003, and you don’t have a Search Engine Optimization (SEO) Strategy to figure out a way to feature amongst the top few search results on Google for keywords that your users would be using? That would be the equivalent of showing up at a football game without shoes.

Cut forward to 2026, if the Agent Economy is the equivalent of Google in 1997, then very soon, if you do not have an Agent Discovery Optimization (ADO) strategy, you will be the one showing up at the football game without shoes! Let me elaborate.

“I have been spending the past few weeks experimenting with OpenClaw and trying to form my mind map on how such Agents will evolve and how they can impact users, businesses, industries, and economies. This morning, I chanced upon a tweet from @andrewchen of a16z, where he spoke about new internet products being designed Agent-first as opposed to Human-first, with CLI endpoints and clean documentation becoming the default callable primitive for other agents — and actually coined the term ‘Agent SEO’ for the process by which service endpoints work to become discoverable in this agent-first world.”

This is an exciting prospect and resonated strongly with the way my thinking has been evolving on the impact of agents on marketplaces, the diminishing relevance of the matching problem as agents eliminate search and discovery friction and the implications for platform ecosystems where breadth of the ecosystem in terms of a large variety of complementors may not quite have the same switching cost effect as the depth of how the platform understands the user and the context (since memory is the superpower in the agentic world).

I foresee agentic marketplaces where service endpoints compete with each other to be the preferred option for API/tool calls, and the decisioning criteria for agents on which endpoint to pick could be very different from the way humans would think (assuming human biases are not predominant in the agentic decisioning process — instead, it is the pure goal-seeking objective set for the agent). In this world, since human emotions are mediated by the agentic process, aspects such as brand, UI, experience, etc., would not be relevant. Instead, the agent would choose which endpoint to utilise based on clear objective criteria, forming the basis of the new world of Agent Discovery Optimization.

The Five Layers of the ADO stack

Agent Discovery Optimization (ADO) would be a multi-layered construct covering aspects such as: can the agent find you, how can the agent discover what you can do, how easy is it to get started, how smooth is the experience, and how effective is the outcome?

Discoverability is the baseline, entry ticket. Having standardized schemas, machine-readable documentation, and easy agent-friendly onboarding will become mandatory. Developers would need to assume that machine-readable documentation is a must-have, along with a human-friendly version. Harness built harness_describe so agents can ask which tools are available and receive a list in return. No docs to read, no hardcoded list. Just ask and find out.

Capability encompasses the breadth of functionality exposed by the service to other Agents. Service endpoints that expose a variety of features will be preferred over those that have a limited scope. For example, an e-commerce service where only search is exposed would limit agents designed for commerce tasks, and a personal finance service where only the current stock price is exposed, and no historical data is served, would be a limiting factor. Google Workspace CLI exposes the full suite. Gmail, Drive, Calendar, Docs. An agent can do everything a human can. That’s breadth.

Onboarding is the first step in the flow where real friction comes to the fore. If the point of technological progress is to reduce friction, then services that let agents sign up, set up a sandbox, authenticate with OAuth, and start with working samples, without jumping through hoops or getting stuck in error loops, will win. Humans have limited patience. Agents have zero patience. Abandoning one service and trying another is trivial for them. Stripe got this. A developer goes from sign-up to first API call in under five minutes. OAuth, test environment, working code. That’s onboarding.

Reliability requires agents to serve their end-users without failures or errors. These are systems expected to run 24x7, and if current experiences with rate-limiting are any indication, trouble-free operation would become more than a hygiene requirement in the agentic world. Besides, in a real-time world, response latency (especially p99 tail latency) determines whether a service is suited for mission-critical tasks. Twilio handles billions of API calls monthly. When an agent sends 10,000 SMS verification calls in an hour, rate-limits are not theoretical; a 429 response is not an inconvenience, it’s a system design flaw.

Finally, Performance is eventually what agents want to integrate for. Optimizing token usage, ensuring consistent, deterministic output, establishing high standards for output quality, and ensuring reliable, glitch-free operation will increasingly be key factors in determining agentic choice. Alpaca APIs support trading, real-time market data, and a consistent schema — the agents don’t need to guess. An agent can’t manually fix errors. If the API returns garbage every 100th call, the agent’s downstream task fails. That’s performance.

New Game, New Rules

While these five factors would play a significant role in influencing top-of-funnel choice decisions for the agent, continued usage and engagement would be driven by an additional set of higher-order factors, such as: Repeat-caller caching, Context persistence, Progressive feature unlock, Memory compounding, and Loyalty pricing. If ADO drives, which agent chooses which service endpoint for a specific need, ongoing usage will be determined by these higher-order factors.

To sum up, we could be at the threshold of a new world on the internet where the brand as we know it, evolves from the conventional definition of “a set of mental associations, held by the customer, which add to the perceived value of a product or service,” (Keller, 1993) manifested for the consumer in the form of logos, advertising, packaging, experience and so on, to become more technologically manifested in the form of uptime percentages, latency, clarity in documentation and schema and such. The brand’s reputation is not just a perception measured in survey panels, but actual data reflected in the API dashboard.

In the Agentic world, Capabilities are primary, and discovery is not an assumed outcome.

We will have a new game in town — ADO.