Production

Reliability

Production primitives applied at the one chokepoint where actions execute: async tools, retries with backoff, timeouts, and durable checkpoint/resume.

Resilient execution

resilient.py
runtime = (
    Antraft.agent(model, tools, task="...")
    .allow(["fetch"])
    .resilient(retries=2, timeout=30, backoff=0.5)
    .build()
)
  • Sync and async tools both work — sync tools run in a thread so they can't block the loop.
  • Failed calls retry with exponential backoff.
  • Each call is bounded by timeout seconds.
Note
Tool, validation, and RBAC errors are non-fatal: the error is reported back to the agent so the model can adapt, and the run continues.

Checkpoint & resume

Every executed action is checkpointed atomically, so a crashed process can resume exactly where it stopped — including its accumulated budget.

checkpoint.py
# write a checkpoint after each step
runtime = Antraft.agent(model, tools, task="...").checkpoint("run.ckpt").build()
await runtime.run()
 
# later, resume from disk
runtime = Antraft.resume(
    Antraft.agent(model, tools, task="..."),
    "run.ckpt",
)
await runtime.run()

What a checkpoint contains

  • Run context — action count, history, tokens, cost, approvals.
  • Agent state — the full conversation and bookkeeping.
  • Written atomically (temp file + replace) so a crash never corrupts it.
Next →
Observability