The era of throwing infinite subagents at every search query is over, or at least it is at TinyFish. In a recent technical breakdown, the team behind TinyResearch revealed that they stripped out autonomous subagents in favor of deterministic tool calls to fix reliability issues in their deep research system. The core philosophy? Stop letting the LLM improvise where it doesn't need to. By moving planning directly into the orchestrator and treating ranking and extraction as simple functions, they cut costs and latency while boosting factual accuracy.

The Death of the Planner

The first major architectural shift was eliminating the separate planner module. Early versions of TinyResearch used a dedicated planner to generate research strategies, but this created a 'snapshot problem.' The planner would hand off instructions to the orchestrator based on incomplete context, causing drift on follow-up questions. For example, if a user asked to focus on pricing risks in Southeast Asia, the planner would often revert to analyzing the whole market. By merging planning into the orchestrator, the system now plans with full awareness of the investigation's history, ensuring that domain-specific constraints stick through every round of the loop.

Deterministic Tools Over Autonomous Agents

The team also abandoned the instinct to spin up a subagent for every search query. They found that steps like ranking results and extracting claims are predictable jobs, not judgment calls. These were converted into plain tool calls that return distilled data. This change significantly reduced the 'agent tax'β€”the overhead of managing multiple LLM contexts. Instead of agents debating what to read, the orchestrator now directs a pipeline: search wide, rank for credibility based on the specific objective, read live via TinyFetch, and extract claims. This pipeline ensures that the final report is assembled strictly from verified claims, not from the model's internal memory.

Benchmarking the 'No-BS' Approach

The results of this streamlined architecture are visible in public benchmarks. In Deep mode, TinyResearch scored 51.74 on DRACO (graded by GPT-5.2) and 46.42 on DeepResearch Bench II (graded by GPT-5.5). On ResearchRubrics, it achieved 77.10 on compliance and 71.04 on strict compliance. These scores reflect a system that prioritizes factual correctness and source credibility over raw speed. The system enforces a depth budget, typically running 40 searches and reading 80 pages in under 15 minutes, but it refuses to write a report until the question is fully covered by extracted evidence.

Key Takeaways

  • Planning must live inside the orchestrator to maintain context across multi-turn research tasks.
  • Subagents are unnecessary for deterministic tasks like ranking and extraction; use tool calls instead.
  • Source ranking must be objective-aware; a government page is useless for sentiment analysis.
  • Reports should only be generated from extracted claims, not from raw page summaries, to prevent hallucination.

The Bottom Line

If your agent is spending more tokens coordinating subagents than actually finding facts, you are building a bureaucracy, not a brain. TinyResearch proves that a tight, deterministic loop beats a loose swarm of autonomous agents every time.