From Hardware to Harness
dmarz
Session summary
The presenter frames the “harness” as the execution surface where agents interact with markets and tools, and divides it into four areas: identity/authentication, payments and blockchain-style transactions, tool calls, and cooperation. This talk zooms into search as a key tool interface, emphasizing how queries over the open or private internet can leak information and why privacy is central if agents are to participate efficiently in markets. The speaker distinguishes between local and public search, observing that many agent operators now scrape and cache large data sets proactively so agents can query them locally rather than relying on fragile or changing external APIs.
The presenter then surveys agent-targeted search services and private search designs. They describe work inspired by the Tiptoe paper, where both queries and the index are encrypted and a server returns encrypted results that only the querying party can decrypt. They also instrumented an agent to study its web-search behavior and found it issues bursts of parallel searches and then spends time reasoning. From this, the presenter infers that reasoning time, not network latency, often dominates, suggesting agents may tolerate slower round-based anonymity protocols such as mixnets or DC-nets.
Key takeaways
- The presenter divides the agent harness into identity/auth, payments and blockchain-style transactions, tool calls, and cooperation, and focuses here on search as a privacy-sensitive tool interface within that surface.
- Agents increasingly combine local and remote search; many operators pre-scrape and cache data such as patents or daily developer updates so agents can query locally and avoid dependence on unstable public APIs.
- Based on measurements of an instrumented agent, the presenter finds that agents issue parallel bursts of web searches and then spend time reasoning, so reasoning latency often dominates and round-based anonymous communication protocols can be acceptable for search workloads.