The SLO Dashboard

An agent in production carries a pager. This lesson gives yours its four numbers.

Your win for this lesson: a dashboard tracking your agent's four SLOs against targets you set — the rubric's first M4 exercise. Prerequisites: Lesson 0004, Lesson 0012 (the traces are your data source).

1 · Pick an agent you actually run

Use your most-developed agent from this course (the multi-tool ReAct agent from 0008, ideally with the 0013 verifier), or — better — a real Vernant surface you operate. M4 is about ownership; own something concrete.

2 · Define the four SLOs

The agent-specific four (Lesson 0004, from the Google SRE framing): an SLI is what you measure; an SLO is the target. Set each — or explicitly write "measure first, then set" and gather a week of data before committing a number (a vibe is not a target):

SLI (measured)	SLO (target)	Source
Success rate (task correct per your 0012 evals)		eval pass/total
Latency (p50 / p95 wall-clock)		trace timing
Cost per task ($ from tokens)		usage × price
Escalation rate (handed to human / said "uncertain")		run outcomes

3 · Build the dashboard

Aggregate your per-run trace records (Lesson 0012) into a rollup. Keep it honest, not fancy — a script that reads the trace log and emits a table, or a small self-contained HTML page, is enough. It must show, per SLI: current value, target, and whether you're in or out of budget.

# pseudo: read traces -> group by day -> per SLI compute value vs target
success_rate = correct / total
p95_latency  = percentile([r.latency_ms for r in runs], 95)
cost_per_task = sum(r.cost for r in runs) / len(runs)
escalation_rate = escalated / total

Pass: the dashboard renders real numbers from real runs, each against a target, with a clear in-budget / out-of-budget signal. If any SLO is "measure first," the dashboard shows the measurement accumulating toward a target-setting decision.

4 · Make one number move

Change one lever (drop effort, lower k, remove the verifier) and watch the tradeoff surface react: cost, latency, and success rate move together (Lesson 0004 — one tradeoff surface, not independent dials).

Pass: you made a change and can point to at least two SLIs moving in opposite directions — the concrete proof that you tune the system, not the prompt.

Ownership begins here: the agent now has first-class metrics and targets. "Is it working?" is no longer a vibe — it's a dashboard with a budget. Everything in M4 hangs off having these numbers: guardrails protect them, checkpoints trade throughput for safety against them, and the improvement loop moves them.

Self-grade

All four SLOs — success, latency, cost-per-task, escalation, each with a target or explicit measure-first plan
Real data — dashboard reads actual run traces, not hand-entered numbers
Budget signal — each SLI shows in/out of budget against its target
Tradeoff surface felt — one lever moved, ≥2 SLIs reacted in tension