M4 Core Vocabulary

The four clusters of Stage 4 — the ownership stage. Source: the Agentic Operations Knowledge Map. Sister sheets: M1 · M2 · M3.

The M4 gate: you can run agentic systems in production against reliability, cost, and safety targets, and drive a continuous-improvement loop without being told what to fix. Arc signal: own SLOs for a live agent and independently run an observe → fix → verify cycle.

Cluster A — Reliability engineering

SLOs for agents: Success rate, latency, cost-per-task, and escalation rate as first-class metrics. The canonical SLI/SLO/SLA framing: Google SRE Book, ch. 4.
Graceful degradation: Fallbacks, retries with backoff, and circuit breakers for flaky tools.
The reliability intuition ⚠️: Production reliability is an engineering discipline, not a prompting problem.

Cluster B — Cost & performance governance

Cost attribution: Cost per run, per user, per feature; token budgets as a design constraint, not an invoice surprise.
Model routing: Matching task difficulty to model tier instead of defaulting to the biggest model.
Ongoing cost levers: Caching, context compression, and prompt optimization — continuous work, not one-time setup.

Cluster C — Safety & control

Guardrails: Input/output validation, injection defense, and tool-permission scoping (least privilege). Standard risk taxonomy: OWASP Top 10 for LLM Applications — prompt injection is LLM01.
Human-in-the-loop design ⚠️: Deciding which actions require approval and how to surface them. The hard part is placing checkpoints by risk without killing throughput.
Auditability: Every consequential action traceable to a decision and its context.

Cluster D — Continuous improvement & leadership

The improvement loop: Observe → evaluate → hypothesize → change → re-evaluate, run continuously, with human feedback recalibrating the metrics.
Regression suites: Gate every prompt, model, and tool change before it ships.
Team practices: Onboarding, prompt/skill libraries, incident review, and documentation as living assets.

The three critical intuitions

⚠️ The model will change under you. New versions silently shift behavior — your eval suite is your only stable ground truth.
Cost, reliability, and quality are one tradeoff surface. You tune the system, not the prompt.
Ownership means owning the feedback loop, not just the current build.

The three M4 exercises

Define SLOs for one production agent and build a dashboard that tracks success rate, cost, and escalation.
Design and defend a human-in-the-loop checkpoint policy for a high-stakes action, then measure its throughput cost.
Run a full improvement cycle: catch a live regression, ship a gated fix, and prove the eval suite would have blocked the bad version.

Signals of readiness for operational ownership

You reach for a trace before a prompt when something breaks.
You can say "no" to adding an agent or a bigger model, with a cost/reliability argument.
You treat the eval suite as the source of truth when a model version shifts behavior.
You design human checkpoints by risk, balancing safety against throughput — without being asked.
You close your own improvement loops: spotting drift, shipping gated fixes, proving the fix held.