Production-ready AI agents fail differently than conference demos. Stages and investor decks are full of agent demos that work. A clean alert. A fluent root-cause analysis (RCA). A green check. Applause.

Then the same pattern meets a partial dashboard, three services blaming each other, a runbook from 2023, and a human who still owns the pager. The demo did not lie about the model. It lied about the environment.

This post is an umbrella for the failure modes we keep relearning while shipping AI agents for SRE and agent workflows — and the receipts that make production different from theater. It is intentionally a map of linked lessons, not a new architecture essay. Steal the checklist; keep your stack.

The Demo Contract (What Quietly Gets Assumed)

Demos assume:

  • Telemetry is complete and labeled the way the prompt expects
  • The “right” runbook is in context
  • Tool calls finish fast and return tidy JSON
  • The stop condition (“investigation complete”) means something is true
  • A human will not approve blindly under load

Production violates every line. Models stay fluent anyway. That fluency is the hazard.

Failure Modes Worth Naming Out Loud

Each mode follows the same shape: demo illusion, production reality, receipt.

1. Fluent but wrong

The demo illusion: The report looks like an RCA. “The database failed due to high CPU.” Decisive prose. Proves almost nothing.

The production reality: Structural completeness is not proof — the looks-right heuristic. The cure is prove, then narrate: fixed stages emit checkable evidence before the model narrates. See Evidence-Gated RCA (multi-backend RCA — metrics, logs, traces — with gates before summary).

Receipt first (illustrative — not a product schema), narrative second:

{"query_id": "tx_992", "cpu_spike_pct": 98, "blocked_pid": 412, "window": "last_15m"}

The receipt: machine-checkable evidence fields exist before presentation is allowed to speak.

2. Open loops that quit early

The demo illusion: Think → tool → think → “investigation complete.” Green check. Applause.

The production reality: Unconstrained loops are polite quitters — thin skim, “nothing to see,” done. Fixed stages with gates beat vibes. AI incident triage gathers metrics, deploys, and similar incidents before proposing where to look.

Illustrative gate (shape only):

{"stage": "gather", "required": ["primary_identity", "kpi_value"], "passed": false, "reason": "missing_kpi"}

The receipt: stage completion criteria a unit test could fail — empty prose cannot pass.

3. Runbook-as-only-navigation

The demo illusion: Ship a forty-page notebook per failure mode; the agent “follows the runbook.”

The production reality: That is a symptom of no map — topology and verify-first probes, not step 7 of 40. Cross-plane reconciliation (do metrics, logs, and traces agree?) and learn-from-verdict memory belong in the platform, not in another PDF. See Agents Need a Map, Not a Script. Wiki vs executable triage: Beyond Confluence Runbooks (GitOps for what must run; wiki for why).

The receipt: injected estate context at launch; structured probe outcomes — not “branch dispatched…” placeholders.

4. End-to-end whodunits

The demo illusion: Run the full pipeline once on a clean fixture; declare victory.

The production reality: When everything runs at once, every stage looks guilty. Bring up one rail at a time against golden gates under live variance (real production noise, not demo data) — How We Debug Multi-Stage AI Agent Workflows. Score expected vs detected class, not transcript polish.

The receipt: per-stage green rates before you celebrate the final summary.

5. Human-in-the-loop (HITL) theater

The demo illusion: “Humans approve every action” — slide shows a responsible team.

The production reality: Approve-everything creates rubber stamps; approve-nothing blocks the product. Risk tiers beat volume — The HITL Paradox (auto-approve read-only, require approval for mutations, hard-deny the worst).

Illustrative tiering (not a product export):

tools:
  metrics_query:
    approval: auto          # read-only
  kubectl_scale:
    approval: required
    risk_tier: high         # state change
  shell_rm_rf:
    approval: denied        # hard block

The receipt: time-to-decide on approvals trends up as volume drops — people are reading again.

6. Context death mid-incident

The demo illusion: Short, tidy tool responses; the agent “reasons through” the incident.

The production reality: The session did not fail because the model was dumb. A log wall ate the budget. Tokenomics is finish rate × signal fidelity × cost per successful workflow — compress tool output so the smoking gun survives.

Before compression (what the model would have choked on):

{"level":"error","msg":"connection reset","trace_id":"a1b2", ... 3,800 more characters ...}

After compression (what still fits in context):

error_count=847 window=15m top_msg="connection reset" sample_trace=a1b2…

The receipt: the session finishes with the decisive signal still visible after compression.

7. Agents that never improve the org

The demo illusion: “It learns from every incident” — bigger memory, flashier digests.

The production reality: Digests without human-approved materialization are souvenirs. Print → propose → review → change workflows/policies — The Diary Learning Loop (learning is an approved change to the system, not a bigger vector store).

Illustrative proposal record (shape only):

{"pattern": "deny_tool:deploy_prod", "count_7d": 12, "proposal": "attach_policy:deploy_guard", "status": "pending_review"}

The receipt: reviewed proposals per week — not generated paragraphs per week.

A Compact Receipts Checklist

Print this for demos that claim to be “production-ready.”

Claim Ask for the receipt
“We found the root cause” Which evidence keys / identities / KPIs were emitted before the narrative?
“The agent finished” Which structural gate passed? Would a wrong answer with the same English still pass?
“We follow the runbook” Is there a map (topology + probes), or only a linear script?
“We tested it” Can you green stages independently under live variance (real prod noise, not fixtures)?
“Humans are in the loop” What is auto-approved, what needs a person, what is hard-denied?
“Cost is under control” Finish rate and cost per success, not spend per chat.
“It learns” Show an approved change that altered policy or workflow — not a bigger vector store.
“It investigates like an SRE” Does it establish identity and onset before deploy theories? Show ruled-out branches, not one hero narrative.

If the seller cannot produce receipts, you are buying theater seats.

Slide deck outline (internal pitch)

Use this as a four-slide arc for leadership or a sprint review — no conference badge required.

Slide 1 — The hook: Same alert, two paths. Fluent false RCA vs receipts-first investigation. The demo lied about the environment, not the model.

Slide 2 — The hazards: Name four polite failures — looks-right prose, open loops that quit early, runbook-as-only navigation, HITL theater (rubber stamps under load).

Slide 3 — The new standard: Map + verify-first probes, bring-up discipline (one stage green at a time), gated workflows. Link to deep dives; skip the product tour.

Slide 4 — The payoff: Humans keep judgment. Agents earn trust with artifacts — evidence keys, gate passes, approval tiers, compression that preserves the smoking gun.

Elevator version: “We don’t need smarter models first. We need receipts — proof the investigation earned its summary before the channel sees it.”

What receipt do you wish you had asked for before the last agent pilot? Find me on GitHub or LinkedIn.

🚀 We’re building AI-powered SRE at StackGen. If you’re tired of 3 AM pages and want AI agents that triage incidents, run diagnostics, and draft RCA reports — check out ai.stackgen.com and try our new SRE offering.