Developers Are Ditching OpenClaw for an Agent That Actually Gets Smarter

Two months ago, Hermes Agent did not exist. Today it has 64,000 GitHub stars, 4,200 forks, 142 contributors, and developer forums full of “I ditched OpenClaw for Hermes” posts. The timing is not coincidental — Hermes launched in February 2026, right as OpenClaw’s security problems and Anthropic’s OAuth enforcement were pushing developers to look for alternatives. But Hermes is not just an OpenClaw replacement. It is trying to solve a problem OpenClaw never attempted.

The thing that makes it different

Every other agent framework treats skills as something humans write. You figure out a workflow, write it down, and the agent follows it. The skills are only as good as what you put in. Hermes inverts this. When it completes a complex task, it automatically generates a reusable skill document from what it did. Next time a similar task comes up, the agent pulls that skill and runs faster. Over repeated use, the skills refine further. The agent gets measurably better at your specific workflows without you doing anything extra.

Community data puts the improvement at 2-3x execution speed after 10-20 similar tasks. After months of use, agents reportedly stop asking questions they already know the answer to. The compound effect is real and it is why the comparison to OpenClaw often stops at “migration path” and starts at “fundamentally different architecture.”

What shipped this week

Hermes v0.8.0 landed April 8 with 209 merged PRs — the largest release since launch. The headline additions: native Google AI Studio support with models.dev integration for automatic context length detection across providers, live model switching across CLI and all messaging platforms simultaneously, and approval buttons on Slack and Telegram so agents request permission via native platform UI rather than requiring users to type commands. MCP OAuth 2.1 with OSV malware scanning for extension packages also shipped — a direct response to the security incidents driving the migration wave.

The week before, v0.7.0 brought pluggable memory backends, credential rotation, and a Camofox anti-detection browser integration. Two meaningful releases in five days from a framework that is two months old.

Why the timing matters

OpenClaw’s CVE-2026-25253 (CVSS 8.8) — an SSRF vulnerability patched in v2026.4.9 — gave many developers the concrete reason they needed to evaluate alternatives. Hermes has zero agent CVEs. The built-in migration command hermes claw migrate lowers the switching cost further. MiniMax M2.7 is now one of the most-used models inside Hermes, and NousResearch is co-developing future releases with MiniMax specifically for agent workloads — a model provider partnership OpenClaw has never had.

The limitation worth noting: Hermes supports six to seven platforms today versus OpenClaw’s 50+. If your workflow depends on iMessage, LINE, WeChat, or niche enterprise integrations, Hermes is not a full replacement yet. It is also not a coding agent — Claude Code and Cursor still win on software engineering tasks. Hermes is strongest on repetitive knowledge work where the learning loop has time to compound.

What to watch

A separate project called hermes-agent-self-evolution — using DSPy and GEPA to automatically optimize Hermes skills and prompts without human input — is the most significant long-term signal. If autonomous skill optimization matures, the gap between Hermes and frameworks relying on human-authored skills will widen considerably. The self-improving loop is not a roadmap feature. It is already running in a parallel repository.

For developers still on OpenClaw, the practical question is whether the platform integration gap matters for your specific setup. If you are running on the six platforms Hermes supports and your workflows are repetitive enough for the learning loop to activate, the migration math is increasingly favorable.