NEW Hermes Agent Speed Upgrade (Update NOW)

Hermes Agent just got a 4-PR speed overhaul — startup cut from 10–15s to 2–3s, terminal tools now poll at 5ms instead of 200ms.

What changed

Four pull requests landed in Hermes and they stack. The creator calls the combined effect "adaptive context" — the agent stops preloading work it doesn't need and stops sleeping longer than the commands it runs.

PR 28864 — CLI startup

Hermes used to import the full OpenAI SDK on every cold start just to patch one internal class. That cost ~160–170ms and ~30MB of RAM per process, even when you never call OpenAI. Now the OpenAI client loads lazily, the moment a model that needs it is hit. The Windows machine the creator tested went from 10–15s cold start to 2–3s; on Mac it was instant. VPS users running multiple workers will feel this hardest because bloat compounds.

PR 28866 — chat loop cleanup

A 31-turn run with tool calls was hitting ~400,000 Python function calls per conversation. After the cleanup it's ~213,000. Two specific changes: reuse already-loaded config instead of deep-copying it every turn, and cache the per-provider reasoning-pad decision instead of recomputing it. Per-turn saving is small (~1ms for a single chat) — the win is on busy gateways and high-concurrency setups where admin work stops starving LLM wait time.

PR 28957 — lazy compression setup

Every agent construction used to immediately walk the auxiliary provider to ask "can this compression model hold a full context window?" — a ~400ms network call. Most short chats never get near the compression threshold, so the cost was paid for nothing. Now a flag gates the check until compression actually triggers. Short chats start snappier; long chats still get full compression.

PR 28957's sibling — adaptive terminal polling (the favorite)

The terminal tool was polling subprocess completion on a fixed 200ms tick. A 30ms echo/ls/cat would block until the next tick — the fixed sleep was most of the visible latency. New behavior: start at 5ms, back off to 200ms only for long-running commands. The creator flagged this as the change that finally puts Hermes in the same league as Claude Code or Aquilo for feel, if not capability.

Bottom line

Faster cold start, less RAM, leaner chat loop, snappier tools. Update now.