Building a Working Stack

Combine models, harnesses, tools, and operational safeguards into a stack that fits the work.

Starting points

  • A minimal personal agent stack can pair a local model such as Ollama or Jan with a harness, MCP tools, and memory.
  • An always-on personal agent can combine a workspace such as Odysseus or Ajax with a messaging gateway such as OpenClaw or Hermes.
  • A coding stack can combine OpenHands or OpenCode with repository context (DOX) and sandboxed execution.

Practical considerations

  • Incorporate calibrated decision models such as Jev and RLCD for routing, gating, and confidence-based escalation.
  • Define containment and permission models.
  • Understand the cost and token economics of long-running agents.
  • Observe and debug multi-step runs.
  • Update skills and memory without destabilizing the system.
Back to top