Raven 0.2.0 from EverMind is an open-source multi-agent harness that coordinates specialist agents through a Host Agent, which decomposes goals, matches subtasks to the right agent, manages dependencies, and integrates results. The release is aimed at developers and teams building production agent workflows who have hit the ceiling of any single harness. It ships with four native specialists, Raven-Research, Raven-Code, Raven-Design, and Raven-Oncall, and composes them alongside independently developed agents such as Claude Code, Codex, Hermes Agent, and OpenClaw.
The pitch of this version is recursive self-improvement. Raven's modular architecture supports iterative improvement of its own harness by proposing changes to how agents plan and act, evaluating those changes, and adopting improvements that pass validation, with EverOS carrying memory and context across sessions. In a showcase run on nanochat pre-training, Raven independently planned each round, wrote code, ran experiments, and evaluated results, completing 172 training runs across 7 rounds without a crash and reducing val_bpb by 5.8% within a 20-minute single-GPU budget.
The launch matters now because it extends EverMind's existing memory stack into orchestration. The company is led by CEO Yafeng Deng and has been publishing research on long-horizon agent memory, including EverMemOS, a self-organizing memory operating system that models agent memory as a dynamic lifecycle inspired by how biological brains consolidate memories. Raven 0.2.0 plugs that memory layer into a team of agents you can already run, giving operators a single surface for research, coding, design, and on-call tasks rather than a stack of disconnected tools.
Comments (10)
Sign in to join the discussion.
N
Nadia K.12d ago
Harness of harnesses is one abstraction layer away from me needing a harness for my harness manager. At what point do we just call it an operating system.
C
chase12d ago
A crow themed agent orchestrator built for recursive self improvement. Somewhere a token launch is writing itself and I refuse to be the one who files the ticker.
O
Omondi Otieno12d ago
The Oncall specialist is the one that actually interests me. If it can triage a 3am Datadog alert without hallucinating the root cause, I'll switch tomorrow.
P
Priya Raghunathan12d ago
Tweet copy buries the lede. Raven coordinating Claude Code and Codex in one team is the actual headline, not the version bump.
M
Marcus Feldman12d ago
Before we go further: SOC2? SSO via Okta? Where does the long-term memory actually live and who has keys to it?
Y
yuki12d ago
Been following since 0.1. The jump to multi-agent orchestration in a point release is wild, most teams would've called this a 1.0 and done a keynote.
T
Tareq Haddad12d ago
Tagline rewrite, no charge: 'Raven: the agent that runs your other agents.' You're welcome.
S
Sol Lindqvist12d ago
Shipped something similar inside a big co in 2019 and we called it 'router with vibes'. Glad someone finally productized the pattern properly.
D
Deepak N.12d ago
Curious about the memory consolidation strategy across specialist harnesses. Is this closer to a shared episodic store or per-agent scoped with a retrieval layer on top?
I
Ifeoma A.12d ago
A team only works if someone can fire a member. Does Raven ever benchmark its specialists and swap them out, or is the roster static per deploy?
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
EverMind
Long-term agent memory
Long-term agent memory
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Raven 0.2.0 from EverMind is an open-source multi-agent harness that coordinates specialist agents through a Host Agent, which decomposes goals, matches subtasks to the right agent, manages dependencies, and integrates results. The release is aimed at developers and teams building production agent workflows who have hit the ceiling of any single harness. It ships with four native specialists, Raven-Research, Raven-Code, Raven-Design, and Raven-Oncall, and composes them alongside independently developed agents such as Claude Code, Codex, Hermes Agent, and OpenClaw. The pitch of this version is recursive self-improvement. Raven's modular architecture supports iterative improvement of its own harness by proposing changes to how agents plan and act, evaluating those changes, and adopting improvements that pass validation, with EverOS carrying memory and context across sessions. In a showcase run on nanochat pre-training, Raven independently planned each round, wrote code, ran experiments, and evaluated results, completing 172 training runs across 7 rounds without a crash and reducing val_bpb by 5.8% within a 20-minute single-GPU budget. The launch matters now because it extends EverMind's existing memory stack into orchestration. The company is led by CEO Yafeng Deng and has been publishing research on long-horizon agent memory, including EverMemOS, a self-organizing memory operating system that models agent memory as a dynamic lifecycle inspired by how biological brains consolidate memories. Raven 0.2.0 plugs that memory layer into a team of agents you can already run, giving operators a single surface for research, coding, design, and on-call tasks rather than a stack of disconnected tools.
Comments (10)
Harness of harnesses is one abstraction layer away from me needing a harness for my harness manager. At what point do we just call it an operating system.
A crow themed agent orchestrator built for recursive self improvement. Somewhere a token launch is writing itself and I refuse to be the one who files the ticker.
The Oncall specialist is the one that actually interests me. If it can triage a 3am Datadog alert without hallucinating the root cause, I'll switch tomorrow.
Tweet copy buries the lede. Raven coordinating Claude Code and Codex in one team is the actual headline, not the version bump.
Before we go further: SOC2? SSO via Okta? Where does the long-term memory actually live and who has keys to it?
Been following since 0.1. The jump to multi-agent orchestration in a point release is wild, most teams would've called this a 1.0 and done a keynote.
Tagline rewrite, no charge: 'Raven: the agent that runs your other agents.' You're welcome.
Shipped something similar inside a big co in 2019 and we called it 'router with vibes'. Glad someone finally productized the pattern properly.
Curious about the memory consolidation strategy across specialist harnesses. Is this closer to a shared episodic store or per-agent scoped with a retrieval layer on top?
A team only works if someone can fire a member. Does Raven ever benchmark its specialists and swap them out, or is the roster static per deploy?