EverMind open-sources multi-agent Host Agent framework Raven, positioned as "the harness of harnesses"

Raven's entry point is explicitly framed as a gap: the current agent ecosystem is fragmented, Claude Code and Codex each have their strengths, but there's no unified layer to orchestrate them. Raven defines itself as a "host agent" that receives complex tasks, generates a task DAG, specifies dependencies, lets parallelizable parts run in parallel, and then dispatches each node to the most suitable agent.

This positioning determines that its relationship with coding agents is additive rather than competitive: it sits above Claude Code and Codex, connecting to third-party agents via ACP, CLI, or OpenAI-compatible APIs, with 13 official presets including Claude Code, Codex, Qwen Code, and Kimi Code. In other words, Raven's value proposition isn't that any single capability is stronger, but that it organizes existing harnesses into a team.

What needs to be distinguished: the framework being open-source and positioned as "the harness of harnesses" are confirmed facts; while the judgment that "the ecosystem is fragmented and lacks a unified orchestration layer" comes from the source's narrative and is a source claim.

Raven has four built-in specialized agents, each with clear responsibility boundaries and corresponding benchmarks: Raven-Research handles autonomous deep research, literature reviews, and structured reports with traceable citations; Raven-Code targets agentic software development; Raven-Design produces visual deliverables like PPTs, posters, charts, and web interfaces; Raven-Oncall targets unattended long-horizon workflow automation.

On the numbers, the README claims Raven-Code tops DataAgentBench with 0.8762 Pass@1; the official line also claims Raven-Oncall outperforms Claude Code on AI4AI and AI4S benchmarks across quality, success rate, and cost. Both of these are official claims only, with specific test conditions and values undisclosed, and no third-party reproduction.

So the more prudent reading is: these scores show the team has offered a testable target, not that Raven has already won on these dimensions.

Among the four built-in agents, Raven-Oncall's scenario is the least like "yet another coding assistant": it targets unattended long-horizon workflow automation, capable of running multi-hour experiments, overnight monitoring, and continuous optimization. The source treats this as the key distinction between Raven and interactive coding tools.

This distinction has practical significance. The value of interactive tools is realized between human turns, while the value of Oncall-type scenarios lies in whether the system can keep advancing, keep verifying, and keep converging after the human leaves. Raven's DAG scheduling, memory, and skill retrieval modules only get truly stress-tested in long-horizon tasks.

But note the boundary: the source only describes the scenario and positioning, without explaining the degree of human intervention in such long-horizon tasks, nor providing comparable stability data.

Recursive self-improvement is the most radical part of the Raven narrative, landing on Evolver: it splits the agent main loop into four decoupled strategy modules—Memory (what context is seen this round), Planning (what method to use), Capability (which tools are available each round), Action (what to do next and judgment before execution).

An experimental Curator rewrites the actual code of these four blocks, not just prompts, but also tools, skills, and judgment logic. The key is the constraint: candidate changes must first beat the baseline on a benchmark and pass a validation gate before being installed, and feedback strips out reference answers to prevent "cheating." On first use, Curator also assembles a customized Persona team based on the user's task description, consisting of a lead role plus several expert agents.

This gatekeeping design is the most worthy part of the entire material to take seriously, because it turns "self-improvement" from a slogan into a process with admission conditions. But the degree of autonomy and safety boundaries of Curator rewriting code are not detailed by the source.

EverOS Memory provides three layers of cross-session memory: user context, the agent's own experience, and world knowledge. SkillForge unifies retrieval across local skill libraries, memory, and roughly 114,000 Skills on SkillHub. Proactivity provides event listening and scheduled dispatch for reminders and follow-ups.

Together these three explain a question: why Raven dares to stretch tasks to hours or even overnight. Without cross-session memory, long-horizon tasks would have to rebuild context every round; without unified skill retrieval, agents can't reuse existing capabilities; without event and scheduling mechanisms, unattended operation is out of the question.

It should be noted that the 114,000 figure comes from the source's description and has not been independently verified.

The THRESHOLD project is used to demonstrate long-horizon multi-agent collaboration: about 4 days, 42 rounds of planning/development/validation cycles, fully autonomously producing a Godot 4 FPS game, plus posters, a presentation, and a website.

The RSI experiment is used to demonstrate self-improvement: 7 rounds totaling 172 nanochat training runs with no crashes throughout, reducing val_bpb by 5.8% under a fixed 20-minute single-GPU budget. The source also mentions CFD (dam-break problem overshoot reduced by three orders of magnitude) and FEA ultimate load search optimization cases.

There's also a "self-proving" detail: the release materials—a browser physics mini-game, a 16-page deck, the README itself—were all generated by Raven itself. These showcases are compelling, but all lack one key piece of information: output quality and degree of human intervention. After 4 days and 42 rounds, how much did humans change, and what were the acceptance criteria? The source doesn't say; whether the 5.8% val_bpb improvement generalizes to other tasks is equally unknown.

Look at the evidence in layers: the framework being open-source, its positioning, DAG scheduling, the four built-in agents, the Evolver and Curator gatekeeping mechanism, third-party integration methods—these are the source's descriptions of its own system, relatively credible but still source claims; while DataAgentBench 0.8762 Pass@1 topping the list, AI4AI/AI4S outperforming Claude Code on three metrics, the RSI experiment's 5.8% val_bpb reduction, and THRESHOLD's 4 days and 42 rounds all come from the official README or a single relay post, with no third-party independent verification.

There's also a weaker piece of evidence: a relay claims Raven builds a DAG from prompts, assigns each node to coding agents like Claude or Codex, collects outputs and feeds them back to the scheduler, allowing the DAG to automatically prune or expand. This description aligns with the official narrative's direction, but it's secondhand relay and can only be treated as a source claim.

Additionally, Raven 0.2.0's version status and subsequent roadmap are not clear in the source. Overall, what's worth watching about Raven is its orchestration abstraction and Curator gatekeeping design, not the numbers it reports about itself.