I spent a rainy Saturday reading all seven Hermes Agent release notes in one sitting. It is the kind of weekend activity that sounds boring in the telling and is actually extremely fun if you are the sort of person who enjoys watching a project figure itself out in public. By the end I had a wall of sticky notes, four cups of coffee, and a pretty clear picture of the shape of what had happened.
Between the first public tag on March 12, 2026 and the v0.8.0 release on April 8, Hermes Agent shipped seven numbered releases in twenty-seven days. That is a release every four days, on average. The PR count for those releases, added up, runs into the four digits. The number of contributors grew from sixty-three at launch to well past two hundred.
Those numbers are not the interesting part. The interesting part is that the releases do not look like one long undifferentiated stream of PRs. They sort themselves into four clear phases, and you can see the project change what it is paying attention to every week or so.
Phase 1: Foundation (v0.2.0)
v0.2.0 on March 12 is the public launch, and its job was to ship a working skeleton: the multi-platform messaging gateway (Telegram, Discord, Slack, WhatsApp, Signal, IMAP/SMTP, Home Assistant in one process), a native Model Context Protocol client, a skills system with seventy-plus bundled skills, a centralized provider router with one call_llm() entry point, and git worktree isolation plus filesystem checkpoints as the safety net for an agent that is actually allowed to modify your machine. ACP integration with VS Code, Zed, and JetBrains made it not-just-a-terminal-thing on day one.
This is the "here is what the thing is" release. Everything that comes after is built on top of these five decisions.
Phase 2: Breadth (v0.3.0 – v0.5.0)
The next three releases, between March 17 and March 28, were about extending the surface area in every direction.
v0.3.0 on March 17 added streaming throughout the agent loop, plugin system hooks, and the first of the big memory integrations — Honcho as a memory provider. This is the release that turned Hermes from "one process with tools" into "one process with a living plugin ecosystem and a memory layer."
v0.4.0 on March 23 did platform expansion: WhatsApp Business API, Signal with full attachment support, and a handful of smaller gateway adapters. More front doors for the one agent.
v0.5.0 on March 28 was a hardening release. Concurrency fixes, session race conditions, tool-result handling, provider quirks. The kind of work that does not land as a highlight reel but without which nothing above it works.
Reading these three together, you can see the project trying to answer a question: "now that we have a core, how much of the real world can we reach from it, without breaking the core in the process?" The answer, by the end of v0.5.0, was most of it.
Phase 3: Durability (v0.6.0 – v0.7.0)
Then the focus shifts. v0.6.0 on March 30 and v0.7.0 on April 3 are about making the thing survive.
v0.6.0 added Profiles — multi-instance Hermes, where one installation can run several fully isolated agents with their own config, memory, sessions, skills, and gateway services. It also shipped MCP server mode, so Hermes can expose itself to other MCP clients like Claude Desktop or Cursor, and an official Docker container. And it introduced ordered fallback provider chains, which is where the "swap providers without rebuilding" story starts getting teeth. Two brand-new messaging platforms, Feishu/Lark and WeCom, joined the gateway.
v0.7.0, the resilience release, is where the architecture got genuinely defensive. Pluggable memory providers — memory becomes a provider ABC that third parties can implement, with Honcho as the reference plugin. Same-provider credential pools with thread-safe least-used rotation and 401 failover. Camofox anti-detection browser backend for stealth web work. Inline diff previews for file-write and patch operations. API server session continuity via X-Hermes-Session-Id headers. A security pass against secret exfiltration, with LLM responses scanned for base64 and URL-encoded credentials.
By the end of v0.7.0, the project had stopped looking like a new thing and started looking like infrastructure. The kind you would run under a cron job and not worry about.
Phase 4: Intelligence (v0.8.0)
Which brings us to April 8 and v0.8.0, the release I wrote about in the previous two posts. The headline is the self-optimized GPT/Codex tool-use guidance loop — the agent diagnosing and patching its own failure modes on OpenAI models through automated behavioral benchmarking. But read in the context of the four-phase arc, it is doing something specific: it is the first release where the project turned its attention back inward, on the agent's own reasoning quality, after three weeks of reaching outward. Live /model switching, free Gemini, free MiMo v2 Pro, background task notifications, inactivity timeouts, approval buttons, MCP OAuth 2.1 PKCE, OSV malware scanning for MCP extensions. 209 PRs. 82 resolved issues. Five days after v0.7.0.
What the rhythm tells you
Looking at this as one continuous arc, three things stand out.
The releases have themes, and the themes do not repeat. Foundation, breadth, durability, intelligence. Nobody seems to have decreed that this is how it should go — the project just behaves as if it has a reading of what comes next. That usually means a small number of people are paying very close attention to the whole surface, and everyone else is pulling in the same direction because the direction is obvious.
The PRs come from a lot of hands. This is not one maintainer and six hangers-on. The release notes are studded with handles I do not recognize. Anonymous pull requests from people who showed up last week. The project behaves like a scene, not a codebase. And scenes, when they work, ship a lot faster than teams.
The velocity is not just count — it is compounding. v0.2.0 shipped the router. v0.6.0 shipped fallback chains on top of the router. v0.7.0 shipped credential pools on top of the fallback chains. v0.8.0 shipped live /model switching on top of all three. Each release is not a fresh set of features; it is a layer that assumes the previous release is stable. You cannot do that if the releases are not actually stable. So either the testing is real, or the velocity would have already killed the project. It has not, which tells you something.
It is worth saying that I am not on the Hermes team. I am a fan who reads release notes for fun and runs this site because the project is more interesting than its marketing surface would suggest. What you are looking at, across these twenty-seven days, is seven releases' worth of evidence that agent engineering at the open-source layer got substantially more interesting in March and April of 2026. I do not know what v0.9.0 is going to be. Whatever it is, I am going to read the notes on the day they ship.