CLAW NOVA

Hermes Agent

The autonomous runtime at the heart of CLAW NOVA.

Hermes is not a chat loop. It is an agent runtime. Every CLAW NOVA surface — research, portfolio, workflows, projects — is the same Hermes loop pointed at a different goal.

The loop

plan(goal, context)         → Plan{ steps[] }
for each step in plan:
  if step.tool: call(tool, args) → Observation
  if step.thought: continue
  memory.write(step, observation)
finalize(plan, memory) → Result

Every iteration writes structured memory entries. Reload the page tomorrow and ask the same project a follow-up — Hermes recalls.

Properties

  • Bounded. Plans have a hard step ceiling (default 6) — no runaway loops.
  • Inspectable. Every plan step, tool call, and observation is rendered live.
  • Interruptible. You can pause, edit, and resume mid-plan.
  • Composable. Workflows compile down to Hermes plans. Skills register as tools.

Why not function-calling alone?

Vendor function-calling is one tactic; Hermes uses it where available and falls back to JSON-mode grammars when not. The planner and executor stay separate so we can swap tactics per model without rewriting the loop.

PreviousArchitectureNext Planner