Hermes Agent 0.21.0 Introduces Persistent Multi-Agent Collaboration, Cross-Gateway Direct Messaging, and Runtime Intervention

Nous Research has released Hermes Agent 0.21.0, expanding its existing scheduling and agent execution foundation into a multi-agent desktop collaboration environment with persistent state, observability, and real-time operator intervention. The new desktop app enables Bot Mode by default. Each agent profile can be configured separately with a name, model, memory, skills, and tools, and can receive `@mention`s in shared chat rooms. Meanwhile, `hermes peer` allows agents on different profiles or gateways to exchange direct messages by handle. Conversations are written to a persistent Bot Chat, preserving the records needed for subsequent tracking and auditing.

This milestone builds on the scheduling overhaul introduced in 0.20.5. That release already allowed scheduled jobs to retain memory across runs, skip LLM calls when monitoring scripts detected no new data, and protect against stalled tool loops. At the time, the maintainers also said that complete cleanup and migration guidance would arrive with 0.21.0. Cron agents can now load and update persistent memory, carry output from the previous run into the next with `continuity=true`, and retain a durable scratchpad for each job. Together, these features elevate what were previously isolated capabilities into a shared state layer for long-running, multi-agent workloads.

The new release also strengthens runtime control. `delegate_task` can list active sub-agents, accept real-time steering, and preserve partial results when execution is terminated early. Sub-agent output can be validated against a JSON Schema, and the system can report the cost of each delegation. The default limits have been raised to 250 iterations and 10 concurrent sub-agents, giving operators more room to handle long-running tasks while also requiring renewed validation of cancellation semantics, resource limits, and fault-propagation behavior.

Version 0.21.0 also adds built-in browser control to the desktop app, MCP health checks, server schema token estimates, 30-day usage reporting, and `hermes://` installation links that require confirmation before opening. To improve security and observability, writes to instruction files such as `AGENTS.md`, skills, and memory now require approval, while secret redaction has been expanded across errors, checkpoints, and ACP logs.

For engineering teams, this is more than a desktop interface update. Cross-gateway direct messaging, persistent memory, and browser control create new paths for permission propagation. Before upgrading, teams should test agent identity, memory isolation, job idempotency after restarts, and termination mechanisms. No official benchmarks are yet available for multi-agent success rates, latency, or fault propagation, so deployment costs and reliability must still be measured in each organization’s own environment.