Concepts

Why Agent Reliability Is Different

Traditional services fail in predictable ways: crashes, timeouts, out-of-memory. You monitor latency, error rate, uptime.

AI agents fail differently:

Traditional Service	AI Agent
Crash → HTTP 500	Hallucinate → HTTP 200 (looks fine!)
Predictable failures	Same input, different output
Stack trace tells you why	No stack trace for "wrong answer"
Retry usually works	Retry might give same wrong answer
Infrastructure problem	Reasoning problem

Your APM dashboard says "everything is green" while your agent approves a fraudulent transaction.

Agent-Specific SLIs

Agent-SRE introduces SLIs that measure agent behavior, not infrastructure:

Task Success Rate

Did the agent complete the task correctly?

# Infrastructure SLI: "Did the HTTP request succeed?"
# Agent SLI: "Did the agent give the right answer?"

success = TaskSuccessRate(target=0.95, window="24h")
success.record_task(success=True)  # Based on your evaluation logic

Hallucination Rate

How often does the agent generate incorrect or fabricated information?

hallucination = HallucinationRate(target=0.05, window="24h")
hallucination.record_evaluation(hallucinated=True, confidence=0.9)

Cost Per Task

How much does each agent task cost in LLM inference?

cost = CostPerTask(target_usd=0.50, window="24h")
cost.record_cost(cost_usd=0.35)

Policy Compliance

Is the agent following governance rules?

compliance = PolicyCompliance(target=1.0, window="24h")
compliance.record_check(compliant=True)

Error Budgets for Non-Deterministic Systems

Traditional error budgets assume deterministic systems: a request either succeeds or fails. Agent error budgets account for uncertainty:

budget = ErrorBudget(
    total=0.05,              # 5% error budget (95% reliability target)
    burn_rate_alert=2.0,     # Alert if burning 2x faster than sustainable
    burn_rate_critical=10.0, # Page if burning 10x faster
)

Burn rate is how fast you're consuming your error budget. A burn rate of 1.0 means you'll exhaust the budget exactly at the end of the window. A burn rate of 10.0 means you'll exhaust it in 1/10th the time — something is wrong.

The Agent-SRE Stack

Your Agent Code
    ↓ instrument with SLIs
Agent-SRE (this library)
    ├── SLO Engine        — "Is my agent reliable enough?"
    ├── Cost Guard        — "Is my agent too expensive?"
    ├── Staged Rollout    — "Is agent v2 better than v1?"
    ├── Chaos Engine      — "What happens when tools fail?"
    ├── Replay Engine     — "What exactly happened 3 hours ago?"
    └── Incident Detector — "Something went wrong, what do we do?"
    ↓ integrates with
Agent-OS (kernel)       — Deterministic policy enforcement
AgentMesh (trust)       — Cross-agent identity and trust scoring

Agent-SRE sits between your agent code and the infrastructure. It answers the question: "Is my agent making good decisions?" — not "Is my server up?"