On-Call When the Model Is in the Path
James Okafor · · 9 min read
The On-Call Assistant is generally available as of this week. I want to be precise about what that means, because "assistant" is a word vendors use to hide an unbounded agent.
Ours is a governed workflow on top of the same inference pipeline as everything else. It can read the observability stack the engineer already has access to. It can draft a mitigation. It cannot take a mutating action that we classified as high-risk.
What it can do
- Query traces by
request_id, workflow, region, and error code - Summarize error clusters over a time window
- Compare p95 latency and cache hit rate against the seven-day baseline
- Pull the last signed records for a failing template
- Draft a customer-facing incident paragraph from the structured timeline
- Suggest the next runbook step, with links to the runbook IDs we published
Those are reads and drafts. They are logged. They inherit PII redaction from the project. An on-call engineer in Johannesburg does not get a raw prompt from a Dubai legal workspace because they are holding the pager.
What it cannot do
- Mute, acknowledge, or rewrite an alert rule
- Edit a routing chain or residency pin
- Page a provider or open a vendor ticket
- Raise a cost envelope
- Resume a dual-controlled legal or risk run
- Disable cache, PII, or sampling on a production project
Those stay on dual control, or they stay on the human's existing tools with their existing IAM. We did not build a second, weaker IAM inside a chat box.
Why this shipped now
We have had twelve enterprise tenants on the pipeline long enough to know the 2 a.m. questions. They are not "what is an LLM." They are "is this a provider 503, a pin fail, or a template that started attaching the whole corpus." The assistant is a faster index over answers we already record.
If the assistant is down, on-call still works. Dashboards, exports, and runbooks do not depend on it. That is a product requirement, not a slogan. A tool that you need in order to see the outage is part of the outage.
James Okafor
VP of Engineering
James Okafor is Vice President of Engineering at Eridian. He holds the production path of the operating system, from model access to signed record. Routing, evaluation, residency, and the controls that keep institutional workloads inside policy sit under his function. His work is to keep Eridian reliable, reviewable, and fit for regulated use at institutional scale.
