← tutorials.pro

Agent Workflows That Survive Contact With Reality

Agents · updated Nov 2026

Multi-agent demos are easy. Multi-agent systems that run unattended for a week are a different sport. This is the field guide to the difference.

Where agents actually pay off

Where they collapse

The verification habits that keep them honest

  1. Every agent turn ends with a probe: a command output, an HTTP status, a file hash. "I think it's done" is not done.
  2. Write-ahead checkpoints: persist state to disk before risky operations; sessions die, disks don't.
  3. One writer per resource: if two agents might touch the same DB/file/service, assign ownership explicitly.
  4. Human gates on irreversible actions: deletes, spend, outbound messages. Everything else can be pre-approved standing orders.
  5. Kill criteria: decide in advance what "this approach is failing" looks like (N consecutive zero-progress turns, repeated apologies, cost threshold) and enforce it mechanically.

A minimal architecture that works

planner (human or strong model)
  -> worker agents (bounded tasks, isolated state)
  -> verifier (deterministic checks: tests, diffs, probes)
  -> ledger (append-only log of what was claimed vs proven)

The ledger is the part everyone skips and the part that prevents the worst failure: an agent reporting success it never verified.

The honest summary

Agents are force multipliers for bounded, verifiable work and elaborate failure generators for everything else. Design the verification first, then add the agents.