Guides

Durable agents

Long-running goals with checkpoints, memory, HITL stalls, and scheduled wakes. Resume after crash or days later. agent_id: persist — showcase for durable execution, not a product wedge.

Scenario

A multi-day goal crashes mid-checkpoint. On resume, poisoned memory steers the agent wrong while sticky routing keeps the same model mid-task. You need resume spans, memory-poison drift, and idempotent wakes.

What Inferix gives you

TraceForge long-lived traces, resume spans, HITL/wake stages
LensAI cost over multi-session tasks (not single-shot)
RouteIQ sticky mid-task routing (R14)
DriftWatch memory-poison / stale-memory slices
FineForge memory-policy prompts + checkpoint schema versions

Tools & intents

Tools	Intents
state.load/save_checkpoint, memory.retrieve/write, workflow.resume, hitl.request_approval, scheduler.wake, domain.act	multi_day_goal, resume_after_crash, hitl_approval_wait, scheduled_wakeup, memory_grounded_continue, checkpoint_replay

Setup

routing:
  sticky:
    mid_task: true
    key: [tenant_id, task_id]
  rules:
    - name: persist-default
      when: { agent: persist }
      model: owned/general-llm
      fallback: provider/strong

observe:
  traces: true
  # resume span must link new session → original trace_id / task_id

drift:
  slices: [memory_poison, stale_memory, post_resume]
  golden_set: /var/lib/inferix/golden/persist.jsonl

Watch & failures

Orphan spans after resume (T16) — fix baggage propagation
HITL stall mistaken for hang — tag exclusive waits (T20)
poison_memory / stale_memory chaos → DriftWatch
drop_wake — scheduler.wake must be idempotent
crash_mid_checkpoint — no duplicate side effects on resume

Glossary

Checkpoint, durable execution, wake, resume span — see Core concepts.

Owned models →