Editorial illustration for TinyFish Moves Planning Into Orchestrator's Unified Reasoning Loop
TinyFish Fixes AI Agent Planning in Production
TinyFish Moves Planning Into Orchestrator's Unified Reasoning Loop
A search agent that aces a demo can still fall apart in production the moment retrieval comes back incomplete, retries loop on the same failed action, or a task blows past its time and cost budget. TinyFish, a company building web infrastructure for AI agents, published two technical guides this week that confront exactly that gap. One lays out a production architecture for agentic search. The other pulls lessons from building a deep research system that has to run reliably outside controlled test conditions.
The core problem both documents address is the split between what a model reasons about and what the surrounding application has to execute, track, and validate. An agent can decide what evidence it needs, but something else has to manage state, call tools, check outputs, and enforce runtime limits so the whole process doesn't stall or spiral in cost. TinyFish built its system around four web primitives, TinySearch, TinyFetch, TinyBrowser, and TinyAgent, run through a seven-phase control loop. On OpenBenchmarks, the approach hit a 79.0% completion rate on coding tasks at a median cost of $0.050 per task, the lowest in the comparison set.
TinyFish introduced its agentic search architecture and key engineering lessons from building deep research systems. Built on four web primitives (TinySearch, TinyFetch, TinyBrowser, and TinyAgent), the system operates via a seven-phase control loop around the base model.
Why this matters
TinyFish's two write-ups read less like marketing and more like a postmortem, which is what makes them useful. Folding planning into the orchestrator's reasoning loop and cutting the per-query subagent model aren't architectural flourishes, they're admissions that the obvious first design (spin up a subagent for every search) buckles under real traffic. Context-switching overhead, latency, cost creep: these are the boring failure modes that never show up in a demo video but show up fast in a bill.
For anyone building search or research agents right now, the lesson is that the orchestration layer matters more than the model choice. Swapping in a better LLM won't fix an architecture that retries dead-end queries or burns budget juggling subagents that don't need to exist. We'd treat this as a useful checklist against your own stack before you scale, not as a template to copy outright.
TinyFish has obvious incentive to frame its infrastructure as the fix, so the real test is whether other teams report the same bottlenecks once they push agentic search past the demo stage.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Agent4Sci testbed measures agents on 20 scientific ML benchmarks in 11 domains
From Archives Shenfeld et al. compare SFT and SDFT on three tasks for continual learning
Common Questions Answered
What are the four web primitives that make up TinyFish's agentic search architecture?
TinyFish's agentic search architecture is built on four core web primitives: TinySearch, TinyFetch, TinyBrowser, and TinyAgent. These primitives work together within a seven-phase control loop around the base model to enable reliable agent operations in production environments.
Why did TinyFish move planning into the orchestrator's unified reasoning loop?
TinyFish discovered that the obvious first design of spinning up a subagent for every search buckles under real traffic due to context-switching overhead, latency, and cost creep. By folding planning into the orchestrator's reasoning loop, they eliminated these boring failure modes that don't appear in demos but cause production systems to fail.
What production challenges does TinyFish's architecture address for AI agents?
TinyFish's architecture confronts critical production gaps where search agents that perform well in demos fail when retrieval comes back incomplete, retries loop on failed actions, or tasks exceed time and cost budgets. Their technical guides provide solutions for building deep research systems that run reliably outside controlled testing environments.
How does TinyFish's approach differ from typical marketing materials about AI agent architecture?
TinyFish's two technical write-ups read less like marketing and more like a postmortem, offering genuine engineering lessons rather than promotional flourishes. This approach provides practical insights into real-world failure modes and architectural decisions that were driven by actual production constraints rather than theoretical optimization.
Further Reading
- Lessons from Building a Deep Research System - TinyFish
- The Path Ahead for Agentic AI: Challenges and Opportunities - arXiv
- Agentic workflows in production: Mechanisms, costs and trade-offs - CoreWeave
- From Language Models to Agentic AI: A Survey of Autonomous, Action-Enabled, and Collaborative LLM Agents - Cognitive Computation
- Agentic AI-RAN Empowering Synergetic Sensing, Communication, Computing, and Control - arXiv
