I woke up on the morning of April 8, 2026, pulled the latest Hermes Agent build, and halfway through my first session it did something I had not seen before. I had asked it to hunt through a pile of logs for a specific error pattern. It ran grep, got nothing, paused for a second, and said: "That pattern is not in the raw output — I am going to try it on the decompressed archives instead." Then it was right.
I do not think I will ever get completely used to that. But the thing is that a lot of what makes v0.8.0 feel like a different release than v0.7.0 is exactly that kind of small, invisible lift in how the agent reasons about its own work. The release notes call it the intelligence release. That is not marketing language; it is describing what actually changed.
The headline: Hermes patched its own tool-use guidance
The most interesting PR in v0.8.0 is #6120, titled "Self-optimized GPT/Codex tool-use guidance via automated behavioral benchmarking." The short version of what it does is: the project built an automated benchmarking loop that probed GPT and Codex models, identified five specific failure modes in how they called tools, generated targeted guidance strings to fix those failures, measured the improvement, and shipped the result as part of the system prompt. The agent diagnosed and patched itself, with humans in the loop as reviewers rather than as prompt engineers.
Alongside that, v0.8.0 added execution discipline guidance in system prompts (#5414) and thinking-only prefill continuation for structured reasoning responses (#5931). If you are using Hermes with GPT-5 or Codex, the agent you woke up to on April 8 is meaningfully smarter than the one you put to bed on April 7. That is a weird sentence to write.
Live model switching across every platform
The second headline is the /model command (#5181, #5742). You can now switch model and provider mid-session from the CLI, Telegram, Discord, Slack, or any gateway platform. The resolver is aggregator-aware: if you are on OpenRouter or Nous Portal and the model is available there, it keeps you on the aggregator; if not, it hops to the direct provider. Telegram and Discord get interactive pickers with inline buttons — you tap the model you want, you do not type its name.
Pair this with v0.7.0 credential pools and v0.6.0 fallback chains and the provider-agnostic story from the previous post is now complete: you can build, break, and rebuild your model lineup without ever restarting.
Free Gemini, free MiMo
Two big provider adds: native Google AI Studio (Gemini) as a first-class provider (#5577), with automatic integration into the models.dev registry for real-time context length detection, and free-tier Xiaomi MiMo v2 Pro on Nous Portal (#6018) for auxiliary tasks like compression, vision, and summarization. If you were holding off on Hermes because your wallet was tired, the free tier got meaningfully more useful this week.
Background task notifications
notify_on_complete is the kind of small change that reshapes how you use a terminal. Background tasks now automatically notify the agent when they finish. Kick off a training run, a test suite, a build, or a deployment, and the agent goes off and does something else. When the background task finishes, the agent gets a ping and can pick the result up. No polling, no "is it done yet" loops.
Inactivity timeouts, approval buttons, and the rest of the list
Two more changes worth calling out. First, inactivity-based agent timeouts (#5389, #5440): gateway and cron timeouts now track actual tool activity instead of wall-clock time. A long-running task that is genuinely working will never be killed — only truly idle sessions time out. This fixes the single most annoying behavior of cron-based agents, which was that they would get the shovel halfway through real work.
Second, approval buttons on Slack and Telegram (#5890, #5975). Dangerous command approvals no longer require typing /approve — you tap a native platform button. Slack also gets thread context preservation; Telegram gets emoji reactions for approval status.
And a grab bag of other things that each deserve their own post: MCP OAuth 2.1 PKCE authentication (#5420), OSV malware scanning of MCP extension packages (#5305), Matrix tier-1 parity with reactions and read receipts (#5275), centralized structured logging to ~/.hermes/logs/ with a hermes logs command (#5430), config structure validation at startup (#5426), and a security hardening pass across SSRF, timing attacks, tar traversal, and credential leakage (#5944, #5613).
By the numbers
209 merged PRs. 82 resolved issues. One release. Five days after v0.7.0.
If you read that number and it feels wrong — 209 is a lot of change for a five-day window — you are reading it correctly. That is the point. Something about how this project is built is making velocity of that kind sustainable, and at some point we are going to have to talk about why. But that is the topic of the next post.