Myrmex Hive’s security model and distribution were solid. What was thin was the part operators actually touch every day: how do I get value out of this without being afraid of it? This release is about that — a documented golden path, plus seven improvements that make the LLM orchestration loop safer to trust and broader in reach.

Start here: the Golden Path

The most common question from new operators is blunt: can the agents change my hosts, or only look? The honest answer — agents can start, stop, and restart services and run allowlisted commands, but there is no arbitrary file write, and every mutation clears six independent gates (RBAC → risk tier → approval → rate limit → allowlist → signed audit), each failing closed.

The new Golden Path guide lays this out as a capability model and a staged rollout: enroll an agent, get value from read-only tools first, grant one controlled mutation behind approval, then let the LLM drive remediation. Capability is something you grant on purpose, never a default.

Preview before you act: ask --plan

Trust is easier to give when you can see what will happen first. ask --plan runs the full orchestration loop — the model is consulted at every step — but executes nothing. The response lists the tool calls it would have made:

myrmex ask --plan "Restart nginx on agent-nginx if it looks wedged"

Happy with the plan? Drop the flag to let it act. It works fleet-wide too.

One prompt, the whole fleet

Orchestration used to target a single agent. Now it can fan out across many and aggregate the summaries — “how’s the whole fleet doing?” in one call:

myrmex call gateway__ask_gemma \
  --arguments '{"prompt":"Report disk usage and flag anything over 85%","all_agents":true}'

Each agent still gets the exact same bounded, gated orchestration; fleet mode is just the single-agent loop in a loop.

Let it run itself: scheduled tasks

Opt-in scheduled_tasks run an orchestration prompt on a timer and route the summary through your existing alert targets — unattended health checks that work between shifts:

"scheduled_tasks": [
  { "name": "nightly-disk-check", "agent_id": "agent-nginx",
    "prompt": "Report disk usage and flag anything over 85%", "interval_seconds": 3600 }
]

Safer by default, and harder to ignore

  • Risk tiers fail closed. Built-in mutating tools (service_control, run_command) now default to a non-read tier even if you never classified them — so a dangerous tool can’t slip past approval gating just because a list wasn’t updated. A guard test enforces that every built-in tool is classified.
  • Approvals page you. A new pending approval now notifies your configured alert targets, so a legitimate mutation can’t quietly expire because nobody was watching the portal.
  • Metrics carry a trend. get_metrics now surfaces recent history when the Gateway is polling, so the model judges health on a trend rather than a single instantaneous sample.
  • A quality gate for the model itself. An opt-in nightly eval drives the real orchestration loop against a live model and grades its output — format, hallucination, convergence, and prompt-injection resistance — so a model swap that makes the assistant “dumber” gets caught automatically.

All of it is opt-in and backward-compatible: an existing config keeps behaving exactly as before. See the Golden Path guide to put it together.