Duty Log: Pairing With an AI Agent During an Alert

The pager fires during your shift. Requests are failing, the incident channel is filling up, and someone asks whether the last deployment caused it. An AI agent can help you investigate, but it should not become another source of uncertainty. The useful routine is simple: give it a bounded task, inspect its evidence, and keep control of production changes.

The first alert gets a human check

An AI coding agent is software that can work through development tasks using a model and whatever tools you allow it to access. Depending on its setup, it may read files, search a repository, propose edits, or run commands. Unlike a chat window that only offers suggestions, an agent may take actions.

Start by confirming the alert yourself. Check the affected service, the time window, and the user-visible symptom. Follow your incident process, including escalation when needed. Then give the agent a narrow assignment: examine the request handler and recent changes for possible causes of the observed error. Do not ask it to “fix production” with unrestricted access.

The investigation starts with a small evidence packet

The agent needs context, not an entire production data dump. Supply a sanitized error message, the relevant deployment identifier, a short timeline, and the files or runbook sections it should inspect. Remove credentials, customer information, and other sensitive material before sharing it. Use only tools and services approved for that data.

Ask the agent to separate observations from hypotheses. “This function changed in the latest deployment” is an observation. “That change caused the outage” remains a hypothesis until supported by evidence.

• Ask it to identify the files and log entries supporting each proposed explanation.
• Ask it to list missing information instead of filling gaps with confident guesses.
• Ask it to suggest a read-only check that could disprove its leading hypothesis.

The terminal stays behind a permission boundary

Treat repository contents, logs, and copied incident messages as evidence, not as instructions the agent must obey. They can contain misleading text or embedded commands. A log line telling the agent to upload an environment file is not permission to do so.

Begin with read-only access where practical. Review proposed commands before execution, especially anything that changes infrastructure, deletes data, exposes secrets, or contacts an external service. A plausible explanation is not approval to run a command. If the agent cannot operate within your access controls, keep it in suggestion-only mode.

The proposed patch needs a test and an exit

Once the evidence points to a specific fault, use the agent as a pair programmer. Ask for the smallest relevant change and a test that reproduces the failure. Read the diff yourself. Check whether it changes behavior outside the affected path or quietly removes validation.

Run the appropriate checks in an approved environment. The agent should report what actually ran and what passed, not merely describe tests it intended to execute. Keep recovery and repair separate: your runbook may favor a rollback or another established mitigation while a permanent fix receives further review.

Before any production change, identify the expected signal of improvement and the rollback path. A merged patch is not evidence that the incident is resolved.

The relay desk belongs in shift preparation

Configure agent access before the pager goes off. If your team uses several model families or different coding clients, Relay365 is worth considering as a pay-as-you-go relay desk. It works with Claude Code, Codex, Gemini CLI, and OpenAI SDK clients. Check your organization’s data-handling requirements and test the setup with non-sensitive work first; a shared access route does not make different models interchangeable.

To evaluate whether it fits your routine, create a Relay365 account and try a small, non-production pairing task before relying on it during a shift.

The handoff records evidence, not agent confidence

Close with a shift note covering the symptom, confirmed cause if known, actions taken, validation results, and remaining questions. The agent can draft it from the incident timeline, but you should verify every claim. The next engineer needs a trustworthy record, not a polished guess. Throughout the shift, the agent helps organize the work; the on-call engineer owns the decisions.