TinyFish
A third-party plugin based on the Model Context Protocol (MCP) designed to extend AI coding frameworks with real browser interaction capabilities, supporting browser automation and web-based operations.
Timeline (last 90 days)
TinyFish's engineering team found that the hierarchical structure of a standalone planner plus per-query subagents brings coordination overhead and information loss in deep research scenarios.
TinyFish's deep research system abandoned the standalone planner and per-query subagents.
Evidence contracts require the agent to produce verifiable evidence at each stage rather than merely giving conclusions.
This design responds to two typical risks when agentic search is put into practice: first, agent outputs lacking traceable evidence, and second, error accumulation leading to loss of control over the task.
Bounded retries set an upper limit on the number of recovery attempts for retrieval failures, preventing the agent from looping endlessly on invalid paths.
TinyFish published two engineering research articles on agentic search architecture, breaking down its design trade-offs from the perspectives of production architecture and deep research systems respectively.
TinyFish's production-grade agentic search architecture is divided into 7 stages, with the core mechanisms being evidence contracts and bounded retries.
For teams building production AI agents, TinyFish's practice offers design principles worth referencing: constrain agent behavior through architecture, support output credibility with evidence chains, and exercise restraint in multi-agent decomposition.
After removing the standalone planner, the system internalized planning capability into the main agent's reasoning process, reducing handoff costs between components.
By making search retrieval free, TinyFish will lower the barrier to entry for agent products at the infrastructure level and put pricing pressure on similar retrieval services.
13 more timeline events
All Facts (20)
TinyFish's deep research system abandoned the standalone planner and per-query subagents.
50%UnverifiedFor teams building production AI agents, TinyFish's practice offers design principles worth referencing: constrain agent behavior through architecture, support output credibility with evidence chains, and exercise restraint in multi-agent decomposition.
50%UnverifiedAfter removing the standalone planner, the system internalized planning capability into the main agent's reasoning process, reducing handoff costs between components.
50%UnverifiedTinyFish's engineering team found that the hierarchical structure of a standalone planner plus per-query subagents brings coordination overhead and information loss in deep research scenarios.
50%UnverifiedTinyFish's production-grade agentic search architecture is divided into 7 stages, with the core mechanisms being evidence contracts and bounded retries.
50%UnverifiedEvidence contracts require the agent to produce verifiable evidence at each stage rather than merely giving conclusions.
50%UnverifiedBounded retries set an upper limit on the number of recovery attempts for retrieval failures, preventing the agent from looping endlessly on invalid paths.
50%UnverifiedThis design responds to two typical risks when agentic search is put into practice: first, agent outputs lacking traceable evidence, and second, error accumulation leading to loss of control over the task.
50%UnverifiedWhether TinyFish can build a long-term advantage with a free search strategy depends on how much added value it can provide in scraping quality, content cleaning and context optimization.
50%UnverifiedThe report did not disclose TinyFish's specific pricing strategy, free quota limits or technical implementation details.
50%UnverifiedBy making search retrieval free, TinyFish will lower the barrier to entry for agent products at the infrastructure level and put pricing pressure on similar retrieval services.
50%UnverifiedTinyFish is a provider of web retrieval infrastructure that recently made its search capability free.
50%UnverifiedAI agents that read web pages bear two types of costs: retrieval fees (per-request billing for search or scraping APIs) and model processing fees (tokens consumed by body content and inference generation costs).
50%UnverifiedTinyFish is a company that builds web agent infrastructure for enterprises, and its core products include capability modules such as Search, Fetch and Browser.
50%UnverifiedTinyFish chose to run its models on Databricks AI Runtime.
50%UnverifiedIn Hugging Face filtering tasks, the more explicit the prompt and rules, the more reliable the results, and it is recommended to verify results using the official API.
50%UnverifiedTinyFish's Web Agent runs on its own Mako model and is billed per step.
50%UnverifiedBy adding an MCP client plugin to the patch file of the desktop profile, pointing it to the TinyFish endpoint, and passing the API key as a header, TinyFish capabilities can be integrated into DeepSeek Harness.
50%UnverifiedTinyFish以"面向AI智能体的Web操作层"（The web operating layer for AI agents）为定位
50%UnverifiedTinyFish published two engineering research articles on agentic search architecture, breaking down its design trade-offs from the perspectives of production architecture and deep research systems respectively.
50%