Systems / Agent runtime
Agent runtime with hard halt
A production pattern for autonomous agents that must act in the world without becoming unstoppable.
Problem
Agents that call tools, spend resources or publish externally need more than a prompt and a chat UI. Without hard stops, approval gates and observability, "autonomy" becomes an incident generator.
Architecture
The reference implementation in our lab is a multi-process agent home:
- Primary loop - KeepAlive worker that evaluates conditions and executes allowed actions
- Deadman - periodic health check that can force safe exits if the loop is unhealthy
- Sweep - orphan and residue cleanup on a longer cadence
- Edge / intel - read-only signal intake, isolated from write paths
- Poster / notifier - sparse outbound communications with rate limits
Configuration, keys and RPC endpoints stay local and gitignored. Runtime state is recoverable from on-disk position or session files after crash.
Control model
- HALT file - presence blocks new risk-taking actions; exit/sell paths remain available
- Approval files - high-ticket or sensitive actions require an explicit operator signal
- Least privilege - tools are narrow; discovery paths cannot snipe protected assets
- Human top-up only - capital introduction remains a human action
What it produces
Structured ledgers, recoverable open state, operator status surfaces, and optional public closed-book exports.
Status
Lab system in live operation for internal experiments. Pattern is portable to non-trading domains (ops agents, research agents, content pipelines) wherever tool use and risk must be gated.