Raven Builds a Composable Network of Evolving Agent Harnesses

Introduction

As language models move from answering isolated questions to executing long-running workflows, the surrounding agent infrastructure becomes increasingly important. Tool use, planning, memory, feedback, and recovery are all part of what Raven calls a “harness.” Raven’s central proposal is not to hand-design one increasingly complex harness for every domain, but to let a system construct, refine, and combine specialized harnesses automatically.

From individual agents to composable units

Raven treats an executable model–harness pair as a modular unit of intelligence. Each unit can be tailored to a particular model, domain, or class of tasks, then orchestrated at a higher level. Instead of asking one agent to master every capability, the system can divide a large objective into subtasks and assign them to agents with more suitable specializations.

The architecture includes several important roles:

• Host Agent: interprets the overall goal, decomposes it, selects specialized agents, coordinates dependencies, and integrates their outputs.
• Host Archive: preserves experience from earlier tasks so that later executions can benefit from accumulated information.
• EverOS: supports the ongoing management of execution context and experience across tasks.
• Skill Forge: turns useful experience into reusable procedures or skills.

This is more than ordinary multi-agent parallelism. The system is designed to evolve the harnesses themselves. A successful workflow can potentially be distilled into a reusable skill and made available to other agents or future tasks.

Why the approach matters

Many agent systems are optimized around one model, one tool chain, or one domain. Such specialization can be effective within a narrow boundary, but cross-domain workflows expose the cost of maintaining separate systems and the limits of a universal harness. Raven addresses this tension through division of labor: specialized units retain domain-specific behavior, while the Host Agent handles coordination across domains.

The paper also develops a theoretical perspective on when composition can expand reliable task coverage under a shared resource budget. This reframes progress in agent systems. The key question is not only whether one agent is stronger, but whether multiple imperfect capabilities can be organized so that their combined coverage exceeds that of any individual unit.

The experience loop is another important component. If task outcomes can be retained and converted into reusable skills, improvement may become a workflow-level process rather than a sequence of isolated prompt adjustments. Over time, the system could accumulate procedures for recurring patterns of work.

At the same time, the available material provides only a high-level summary and a broad claim of improved performance on complex, long-horizon tasks. It does not specify the full benchmark suite, resource budgets, ablations, or performance margins. Questions about coordination overhead, error propagation, retrieval quality, and the reliability of automatically generated harnesses therefore remain open.

Raven ultimately represents a shift from building one “general” agent to building an ecosystem of interoperable agents. Its practical significance will depend on whether harness construction is reliable, whether learned skills transfer across tasks, and whether the benefits of composition outweigh the additional planning and execution costs.

Source: Hugging Face Daily Papers