Harnesses
Hermes Agent 0.21: bot society, cron memory and steerable subagents
The 31 August release turns Hermes from a single assistant into a small society of named agents, gives scheduled jobs a memory, and lets you steer subagents mid-flight. Here is what changes for a small-business deployment.
- Published
- 2 Sep 2026
- Reading
- 9 min
- Class
- harnesses
half-life 60dfrom 2 Sep 2026
Hermes Agent's 0.21.0 release landed on 31 August, and it is a large one. The headline features share a theme: Hermes is moving from an assistant you talk to towards a set of long-lived agents that talk to each other and keep working when you are not there.
What is new
- Bot Mode. Bundled in the desktop app: every agent profile gets a name, a deterministic avatar and a place in a shared roster, so you can run Discord-style group chats with @-mentions.
hermes peer. Any agent can message any other by handle across profiles and gateways, and replies land in a durable Bot Chat you can read afterwards.- Cron with memory. Scheduled jobs now load and update persistent memory like any other run. A continuity option carries one run's output into the next, and monitor-mode jobs skip the model call entirely when nothing has changed.
- Steerable subagents.
delegate_taskcan list running children, correct one mid-flight or stop it early while keeping partial results, validate output against a JSON schema and report cost per delegation. Defaults rose to 250 iterations and ten concurrent children. - MCP command centre. MCP servers and the catalogue on one page, with drag-in import, background health checks and a fleet cost overlay.
- An agent-driven browser in the desktop app, and six new providers plus a wave of new models in the catalogue.
The part that matters for a business
Two of these change the economics of the always-on jobs we described in what to automate first. A monitor-mode cron job that skips the model when nothing has changed turns a poll-every-ten-minutes check from a steady token bill into an almost free one. And cron memory means a recurring summary can say what is new since last time rather than starting from scratch.
The security changes are just as relevant to anyone following the small-business case. Standing instruction files, skills and memory stores are now protected and need write approval, which closes off the most direct way for injected text to rewrite an agent's orders. There is also a broad sweep to redact secrets from terminal errors, environment-file reads, checkpoints and logs, and per-model overrides so you can correct a context window or price without waiting for a release.
Before you update
It is a big release. Back up your profile data, update a non-critical profile first, and give it a few days before moving production; point releases usually follow a release this size. Check the approval prompts after updating — you will now see write approvals you did not before, and that is the feature working.
Bot Mode and peer messaging are the newest and least proven pieces. Try them on something reversible, such as an internal channel, before letting agents message customers. Our Hermes Agent spec sheet has the standing summary, and the setup guide still applies to a fresh install.
Sources
- NousResearch/hermes-agent releases — v0.21.0 notes, 31 August 2026.