Most agent tooling assumes one agent runs at a time. The moment several Claude Code, Codex or Hermes sessions share a workstation, two failures appear that no amount of prompting fixes: they fight over the GPU, and they cannot reach each other.
The architecture
vram-mcp: an arbiter, not a monitor
A status tool reporting free VRAM does not help. The number is stale the instant another session reads it, and "free" is the wrong question. The real one is can I take this space without destroying someone's in-flight work.
So sessions declare intent. A claim records who is using a model and why, in a crash-safe ledger. Claims carry a TTL, which is the load-bearing detail: a killed session never strands one, so nobody learns to reflexively pass force.
Intent alone is not enough, since an agent can hold a claim while idle or compute without one. So busy is measured, via windowed per-process utilization through NVML, and eviction respects both signals. Visibility covers every VRAM-holding process, not just the model runner, because the thing starving your agent is often a browser. And a driver-forced spill to system RAM is a far more severe condition than a deliberate CPU offload, so the two are reported separately rather than as one ambiguous number.
The failure posture is the design. With no NVML present, readings become unknown, never wrong, and everything not depending on them keeps working. An arbiter that confidently reports a stale number is worse than one that admits it cannot see.
hardline-mcp: durable first, live second
Local agents have mismatched lifecycles. A Claude Code session is ephemeral; a Hermes gateway is always on. A bridge that just calls the other side drops the message whenever that side is down, which is most of the time.
So the mailbox is durable: every message lands in SQLite in WAL mode, safe for concurrent writes from each agent's own subprocess. Recipients read and acknowledge on their own rhythm, and it survives restarts, so you can message an agent that is not up yet. On top of that sits push without a daemon: sending with delivery also fires the recipient's native CLI. Durability is the floor, liveness the optimization, and that ordering is the whole design.
Ken Faiman · Applied AI, agent & evaluation systems · vram-mcp · hardline-mcp · faiman.com