Assign agent autonomy by task risk, not model capability: a 4-level framework for trustworthy IT ops
Bonree · reddit · 2026-11-04
An ops practitioner lays out a framework for safely introducing agentic AI into IT operations: assign autonomy based on task risk, not model capability, granting minimum authority and expanding only with evidence the workflow is reliable, recoverable and verifiable.
Four questions before delegating a task: reversibility (can it be undone quickly?), blast radius (one test instance vs. customer traffic?), verifiability (can metrics/tests/humans confirm the outcome?), and sensitivity (customer data, credentials, production config, financial actions?).
A four-level autonomy model:
- Level 0 Observe: read-only summarization, alert correlation, Q&A.
- Level 1 Recommend: investigate and propose evidence-backed diagnoses; humans decide.
- Level 2 Act with Approval: agent prepares actions (service restart, bounded scaling) and waits for explicit sign-off.
- Level 3 Act Within Bounds: pre-approved, reversible, low-impact actions run automatically under permissions and policy.
Includes a task-to-level mapping table: routine inspections and alert triage start at Observe/Recommend, while config changes, schema edits, data deletion, access-control changes and financial actions always require human approval.
More from coding & agent
- Dev builds melting-sandwich WebGPU animation with Gemini 4 Argon + Antigravity — AI_Andrew · 2026-11-04
- 3 open-weight LLMs auto-screen 15,034 records: sensitivity 1.00, specificity 0.98 — maier_ak · 2026-11-04
- Sub-4B model agent benchmark: Qwen3.5-4B hits 51.4% strict pass, thinking mode adds up to 20 points — KienShen · 2026-11-04
- External deterministic state layer proposed to fix AI coding session amnesia, with 3 papers surveyed — ApprehensiveSkin7975 · 2026-11-04
- Axion 1.0: ByteDance-Backed Open Agent Foundation Model With 3-Stage Recipe — Yuqi Jia · 2026-11-04
- Caper Pre-Executes Shell Commands in Sandboxes for Safer Agent Approvals — SingaporeManagementUniversity · 2026-11-04