Causely moves the diagnosis to the detection side of Anthropic's stage-six trigger

The causal model is built before the incident. Causely encodes, per type of entity, which root causes are possible and which observable signals each one produces, both on that entity and on the entities that depend on it ^[12]. Ben Yemini's examples are faulty error handling on an RPC method, a slow consumer on a topic, and a saturated connection pool on a database client ^[13]. A slow consumer shows up as growing queue depth on the topic and rising latency on whatever is waiting for the result; faulty error handling shows up as a high error rate on that method and on every service that calls it ^[14].

At runtime, metrics, alerts and events are mapped to entities and become observed signals, and the model asks which diagnosis best explains the currently active set ^[15]. Because the signatures were written in advance, that step is a lookup against the model and not a search through raw telemetry ^[15]. The same signals on the same topology produce the same diagnosis ^[16].

That property is what the argument leans on, because the playbook closes the detection half to models. Louis Claxton, writing for Anthropic, said "detection stays entirely deterministic, with no model involved." ^[6] A lookup against a catalogue built by hand before the outage satisfies that sentence as written, and that reading is where Yemini's claim rests ^[16].

For the swap to pay off in someone else's cluster, the fault has to be in the catalogue already. His objection to bands is that each one watches a single metric, so you need a detection script for every failure mode you anticipated ^[9]. A causal model needs the root cause and its signature encoded ahead of time too ^[12]. Either way, somebody has to have thought of the failure first. The unit differs: a band is written per metric, a signature per entity type, so one encoded root cause covers every instance of that type ^[12].

The post calls diagnosis the agent's first and most expensive phase ^[7]. Under the playbook, the tier at which the agent diagnoses is read-only, and of the three tiers only one permits a change to production, through a PR into the review gate or a pre-approved runbook ^[4][19]. Handing the session a primary diagnosis and its evidence chain ^[11] therefore shortens work at 2 sigma, where nothing ships.

The evidence is one vendor's post, written by Yemini and first published on causely.ai ^[17]. It follows a setup post that wired Claude managed agents to a local Kubernetes cluster with Grafana and Causely MCP tools, and worked through the authentication pattern for each ^[18]. It does not report a session started from an Issue measured against one started from a band breach: no token counts and no time to a correct diagnosis ^[20].