Sentient ROMA: recursive task-tree agent framework hits 45.6% on Seal-0
Quote280 words
Explains ROMA's recursive task tree (Atomizer, Planner, Executors, Aggregator) with inspectable nodes, and why compounding errors cripple long-horizon agents. Reports a 5x gain over Sentient's previous framework.
Key claims
45.6% on Seal-0, up from 8.9%
81.7% on FRAMES
99% accurate models drop below 91% success after 10 steps
Sentient 5x'd their own agent performance in under a year. Their ROMA framework hits 45.6% on Seal-0's long-horizon benchmark, up from 8.9% with their previous architecture.
TL;DR on a true architectural breakthrough ↓
➤ Modern AI agents hit a wall with multi-step tasks, even 99% accurate models drop below 91% success rate after just 10 steps due to compounding errors
➤ ROMA aims to solve this as a meta-agent framework that orchestrates specialized agents rather than trying to be one monolithic system
➤ Core innovation is a recursive task tree that breaks complex goals into tractable sub-tasks with transparent, inspectable reasoning at every node
➤ Four-stage lifecycle: Atomizer assesses complexity, Planner decomposes into sub-tasks, Executors handle simple actions with appropriate tools, while Aggregator synthesizes results
➤ Example: "How many $350M+ budget films weren't highest-grossing in their year" becomes three parallel sub-tasks: find high-budget films, identify yearly top grossers, compare and count
➤ Fully transparent architecture lets developers inspect inputs/outputs at every node, making debugging trivial compared to black-box systems
➤ Modular design allows swapping different LLMs, specialized tools, or human verification at any point in the execution tree
➤ Achieves SOTA performance: 45.6% on Seal-0 benchmark (beating Kimi-Researcher by 14.1%, doubling Gemini 2.5 Pro)
➤ 81.7% accuracy on DeepMind's FRAMES multi-hop reasoning benchmark, surpassing previous SOTA benchmark results
➤ Represents 5x improvement over @SentientAGI's own previous framework (Open Deep Search), showing rapid internal innovation pace
➤ Solves the architectural "meta-challenge" while letting developers focus on task-specific prompts and tools
➤ Foundational primitive for open-source community to build reliable multi-agent systems that actually work at scale
Check out the full article here ↓
For more on @SentientAGI, check my thread below.
QUOTING @zerokn0wledge_
t.co/XEvWAx65E6
t.co/XEvWAx65E6
What came next · since 31 Oct 2025
Since then
Zero has written 2 more pieces on Sentient since this one.