Faster search and leaner agent tool calls
Search latency and tool definitions both add up. Today’s picks give agents a faster web-retrieval option, a way to fetch tools only when needed, and a safer space to think through an app before editing it.
In today’s edition
- Perplexity launches Fast Search for API agents.
- OpenRouter defers large tool libraries until needed.
- Lovable adds free workspace-level app chats.
Perplexity adds a low-latency search lane for agents
Perplexity’s new Fast Search preset brings its Photon retrieval-and-ranking engine to the Search API, targeting agent workflows where each search call can slow down the whole run.
- Perplexity says 95% of Fast Search results return in 230 ms or less.
- The company reports 160 ms median latency and 68% lower cost per task than its default preset across six internal agent benchmarks, at comparable quality.
- Fast Search spends less compute on ranking than the default preset.
- Perplexity reports tradeoffs on long-tail and broad-coverage queries: relevance was 0.24 points lower and answer availability about 3 percentage points lower in its internal tests.
What it means Use it for latency-sensitive, repeated agent lookups, but benchmark your query mix first because Perplexity reports modest relevance and answer-availability tradeoffs versus its default preset.
Source: perplexity_ai
OpenRouter lets agents search for the right tool
OpenRouter’s Server Tools Marketplace includes Tool Search, which keeps a large function catalog out of every prompt until the model needs a specific capability.
- Tool Search is aimed at libraries with more than roughly 10 tools or 10,000 tokens of definitions.
- Mark function tools as `defer_loading` so they stay out of the initial prompt.
- The model searches deferred tools by pattern, then receives the best matches with their names, descriptions, and arguments.
- Your application still handles the resulting function calls, while OpenRouter performs the matching during the request.
What it means For agents with sprawling action catalogs, defer tool schemas and load only matched definitions, shrinking repeat prompt payloads without rewriting your execution layer.
Source: OpenRouter
Lovable makes project-aware chat free to start
Lovable now offers a daily free chat allowance for workspace conversations, so you can inspect an app, connected tools, and Cloud data before asking the agent to make changes.
- Free, Pro, and Business workspaces get a daily free chat allowance through October 31, 2026.
- Lovable can inspect an app’s code and query its Lovable Cloud database during a conversation.
- You can bring in information from connected tools such as Notion, Granola, and Linear.
- Generating images or video, or handing work to Plan or Build, still consumes credits.
What it means Use the free chat allowance to interrogate code, cloud data, and feedback before committing credits to a plan or build, especially when deciding between competing changes.
Source: lovable.dev