Recursia: Algorithmic Context Management and Execution Engine for Multi-Agent Workflows

Recursia is an innovative execution engine for multi-agent workflows. Its core uses minimal topological read-write subset routing and attention isolation techniques to significantly reduce Time to First Token (TTFT), enable efficient parallel LLM inference, and solve the context inflation problem in multi-agent systems.

With the improvement of LLM capabilities, multi-agent architectures have become popular in scenarios like automated customer service and research assistants, but they face the challenge of context inflation:

• Sharp rise in TTFT: Models need to process longer inputs to generate the first token
• Soaring inference costs: Long contexts increase computing resources and API fees
• Attention dilution: Key information is overwhelmed by massive contexts.

The core of Recursia is algorithmic context management, with key strategies including:

1. Minimal topological read-write subset: Based on the workflow dependency topology, calculate the minimal context set required for each agent, route on demand, and reduce input length.
2. Attention isolation: Physically isolate the context spaces of different agents to ensure the model's attention is focused on information relevant to the current task.

Recursia's architecture includes two core components:

• Context Manager: Builds dependency graphs, dynamically calculates minimal contexts, and maintains state version control.
• Execution Engine: Parallelly routes groups of agents that can run in parallel, aggregates results, and handles errors (retry and recovery). Comparison with traditional frameworks:

  Feature	Traditional Frameworks	Recursia
  Context Strategy	Full transfer	Minimal subset routing
  Attention Management	Shared space	Physical isolation
  Parallelism Granularity	Coarse-grained	Fine-grained topological parallelism
  TTFT Optimization	Limited	Significant reduction
  (Compared with frameworks like LangChain and AutoGen)

Recursia has achieved significant results in reducing TTFT:

• Mathematical analysis: In linear workflows, the k-th agent in traditional methods processes approximately (k-1)×M context, while Recursia reduces it to a constant level (only direct predecessor output).
• Practical significance: In latency-sensitive scenarios like real-time dialogue and interactive programming, reducing TTFT directly improves user experience (instant feedback).

Recursia is particularly suitable for the following scenarios:

1. Complex reasoning chains: Decompose multi-step reasoning into specialized agents, keeping the context concise (e.g., mathematical proofs, logic puzzles).
2. Tool call workflows: Ensure tool nodes only receive necessary parameters and pre-results (e.g., data analysis pipelines, automated operation and maintenance).
3. Multimodal processing: Agents of different modalities work in parallel, with efficient routing of inputs and outputs.

When applying Recursia, the following points should be noted:

• Accuracy of dependency analysis: The calculation of minimal subsets relies on accurate dependency graph modeling; errors may lead to information loss or redundancy.
• State consistency: Ensuring that multiple agents have a consistent understanding of shared states during parallel execution is a challenge in distributed systems.
• Debugging complexity: Streamlining the context improves performance, but increases the difficulty of reconstructing traces when errors occur.

Recursia represents the trend of evolving from functional completeness to performance optimization:

• Insights: Prompt engineering needs to consider minimizing length; architectures need to balance functionality and efficiency; the success of LLM applications depends on underlying system optimization.
• Summary: Recursia provides an innovative solution to the TTFT and cost problems of multi-agent workflows. Although it is in the early stage, its design concept is worth paying attention to, and it is an option for performance optimization strategies in production-level applications.