Hermes Agent shipped version 0.21.0 on August 31 — the Pantheon Release: more than 2,400 pull requests merged since 0.20.0 and a repository past 230,000 GitHub stars. For anyone running it on their own server, the most practical change is not the size of the jump but a specific one: scheduled tasks now remember what they did on the previous run.
Scheduled tasks that remember
An agent that works on its own needs memory between runs; without it, every task starts from zero. Version 0.21.0 solves this in several ways. Cron agents now load and update persistent memory just like an interactive session. The continuity=true flag chains runs together: each run’s output carries into the next, so a monitor can skip what it already reported yesterday. Every job also gets a durable notepad scratchpad for notes between runs.
One detail cuts cost to zero in the most common case: monitor mode. If nothing changed since the previous run, the task does not call the model at all. A disk or certificate watchdog that checks hourly and only alerts on novelty spends no tokens. Task output can also land in a bot’s Bot Chat, where the bot actually replies instead of just posting a report.
Subagents you can steer mid-flight
Delegating work to subagents went from fire-and-wait to live orchestration: you can list running children, send a course correction to one while it runs, or stop it early and keep the partial result. Outputs support JSON-schema validation, so a subagent returning malformed data gets caught before it pollutes the parent agent. Default limits rose to 250 iterations and 10 concurrent children.
Named bots and agent-to-agent conversations
Bot Mode ships built into the desktop app: every agent profile gets a name, an avatar, and a slot in a shared roster, and you can create group chats where several bots and people talk with @-mentions. The new hermes peer command lets any agent message another by handle, across profiles and gateways, from the CLI or inside a conversation. Replies land in each agent’s Bot Chat, so those conversations are durable and inspectable rather than fire-and-forget.
Desktop, browser, and context
The desktop app keeps persistent connections to several gateways at once — a home server, a laptop, and a cloud instance show up as rooms in the same interface. The agent now drives the built-in browser directly — it navigates, clicks, and reads pages — and MCP server management merged into one panel with background health checks and hermes:// deep-link installs. Default context usage dropped by roughly 50%.
What running it on your own server means
Hermes Agent is free software under the MIT license and installs on your own machine: the real cost is the model you connect and the maintenance. Scheduled tasks run in fresh sessions, and a cron job cannot create more cron jobs — a deliberate restriction that prevents runaway scheduling loops. And since every run starts from a clean session, persistent memory helps but does not replace a self-contained prompt: if the task needs context, write it into the task.
If the question is local agent versus a cloud service, we compare cost and privacy of both routes in Local AI vs. ChatGPT. To test the full flow, the first step is cheap: schedule one trivial task — say, a nightly summary of what you saved to Obsidian — with continuity=true, let it run for a week, and then judge.