Conference 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 Aiden-adjacent AI 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

1. Fluent but wrong

The report looks like an RCA. Structural completeness of prose is not proof. We wrote about this as the looks-right heuristic and the cure — prove, then narrate — in Evidence-Gated RCA (fixed stages emit checkable evidence before the model is allowed to narrate).

Demo output: “The database failed due to high CPU.” Looks decisive. Proves almost nothing.

Production receipt first (illustrative shape — not a product schema), then narrative:

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

Receipt: machine-checkable evidence fields before presentation is allowed to speak.

2. Open loops that quit early

Unconstrained think → tool → think loops are polite quitters. Thin skim, “nothing to see,” done. Fixed stages with gates beat vibes — and AI incident triage forces the agent to gather metrics, deploys, and similar incidents before proposing where to look.

Receipt: stage completion criteria that a unit test could fail.

3. Runbook-as-only-navigation

Shipping another forty-page notebook for each failure mode is a symptom that the product has no map. Topology, verify-first probes, cross-plane reconciliation, and learn-from-verdict memory — see Agents Need a Map, Not a Script (inject estate context at launch; scripts are overlays). Wiki vs executable triage is a sibling trade-off in Beyond Confluence Runbooks (GitOps contracts for what must run; wiki for why).

Receipt: injected estate context at launch; structured probe outcomes; not “step 7 of 40.”

4. End-to-end whodunits

When the whole pipeline is wrong, every stage looks guilty. Bring the board up one rail at a time against golden gates — Bring Up Agent Workflows Like Hardware (green each stage under live variance before adding the next). Score expected vs detected class, not how pretty the transcript reads.

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

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

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

Receipt: time-to-decide on approvals trending up as volume drops (people are reading again).

6. Context death mid-incident

The session did not fail because the model was dumb. It failed because a log wall ate the budget. Tokenomics is finish rate × signal fidelity × cost per successful workflow — compress tool walls so the smoking gun survives.

Receipt: sessions that finish with the smoking gun still visible after compression.

7. Agents that never improve the org

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).

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?
“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.

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

How to Pitch This Internally

If you are championing better agent standards to leadership — or presenting at an internal tech talk — keep the arc short. You do not need a conference badge for this conversation to matter:

  1. Hook: same alert, fluent false RCA vs receipts-first path.
  2. Name four polite failures (looks-right, early quit, runbook crutch, approval fatigue).
  3. Show the map and bring-up discipline without a product tour.
  4. Close: humans keep judgment; agents earn trust with artifacts.

That is the sister narrative to reliability-over-intelligence slides — the builder version with scars, usable in a sprint review as much as on a stage.

Acknowledgments. Built with the StackGen Aiden team — the engineers behind the agent runtime and platform this series describes.

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.