Appendix A: Production Challenges, Checklists and Templates

Version: 1.0 Last updated: February 2026

Navigation: Appendix Index | Appendix B →

This appendix combines two concerns that belong together: what blocks agents from scaling, and the checklists that help teams avoid those blockers. The first half catalogs production challenges. The second half provides actionable checklists and templates.

A.1 Why Production is Hard

Chapter 1 frames the core tension. Capability is rising, but enterprise impact is uneven.

METR measures agent capability via "time horizon." Frontier systems reach approximately 50-minute horizons at 50% success. This capability roughly doubles every seven months, with caveats.

OpenAI's GDPval evaluates real-world tasks across 44 occupations in 9 sectors. It ties evaluation to economically valuable deliverables.

Widely circulated claims that 95% of pilots fail should be treated as a signal, not a law. The reliable takeaway is that pilots stall because of workflow fit, ownership, integration, evaluation, and operational constraints. Base model quality is rarely the blocker.

The Berkeley-led Measuring Agents in Production (MAP) study bridges theory to reality. It documents that production agents are typically short, constrained, and checkpointed. Reliability is the top challenge.

A.2 Challenge Categories

A.2.1 Wrong Problem Selection

Teams build agentic systems to mask unclear processes or to replace stable deterministic flows. Success criteria become fuzzy like "be helpful." The agent cannot reliably terminate or verify.

This happens when the agent becomes a proxy for missing product definition, missing process mapping, or missing ownership.

To mitigate this, apply the triage discipline from Chapter 3. Agents suit adaptive multi-step work under uncertainty. They do not suit deterministic execution or simple retrieval. Define "done" as testable outcomes. Then constrain the action space.

A.2.2 Data Readiness

Demos work great on curated documents. Production fails due to stale knowledge bases, mismatched IDs, inconsistent taxonomies, and missing entitlements. RAG returns plausible but wrong evidence. The agent cannot retrieve the needed field to complete a tool call.

Enterprise data is fragmented across systems of record. Semantics are inconsistent. Permissions are complex. Context is a product in itself.

To mitigate this, treat context engineering as an engineering subsystem. Build step-specific context builders, budgets, provenance, and a retrieval test suite for both offline and online use. Invest in entity resolution, canonical IDs, and data contracts before granting autonomy.

A.2.3 Integration Complexity

The system works in staging but breaks on real auth, rate limits, partial failures, network timeouts, and non-idempotent APIs. Wiring tools requires high manual effort to map schemas and manage secrets across many systems.

To mitigate this, standardize tools as contracts. Use typed schemas, deterministic validation, idempotency keys, and safe defaults. Centralize tool access and enforcement where possible. Place policy gates outside the model loop.

A.2.4 Architecture Mistakes

Long-running wandering trajectories emerge. Unexpected tool calls appear. Costs run away. Multi-agent systems get added early. This increases coordination failure without clear returns.

To mitigate this, start with a strong single-agent scaffold. Add agents only when roles are crisp and measured. Keep runs short. Require checkpoints. Force verification and termination criteria. Use MAP's "boring agents win" pattern as the default: constrained steps, frequent human or deterministic gates, and reliability-first design.

A.2.5 Evaluation Gaps

No regression suite exists. No scenario replay exists. No online monitoring ties to task success. Incident response is guesswork. You cannot reconstruct what context the agent saw and why it acted.

To mitigate this, use harnesses, traces, and replay as first-class artifacts. This is the core of Chapters 5 through 8. Adopt online evaluation loops and shadow mode before automation. Track task success, policy compliance, tool error rates, and cost per trajectory. Do not track only model-level metrics.

A.2.6 Governance Gaps

Agents access more systems than humans. They move faster than manual review. They can be tricked into unsafe actions. Policy enforcement exists only in the prompt.

To mitigate this, treat agents like privileged machine identities. Apply least privilege, scoped tokens, audit logs, and separation between plan and act. Use deterministic runtime blocks for forbidden actions. Do not rely on prompt-only policies.

A.2.7 Organizational Bottlenecks

Productivity gains do not compound. Employees feel anxious. Training is insufficient. Workflow redesign does not happen. Shadow AI emerges where employees use personal tools because official agents do not fit workflows.

To mitigate this, redesign workflows around agent capabilities. Do not just insert agents into existing processes. Invest in training, incentive alignment, and feedback loops.

A.2.8 Economics and Procurement

Token costs, tool-call costs, and latency explode when trajectories are long. Procurement and security reviews slow down iteration. Learning loops break.

To mitigate this, make step budgets, retry budgets, context budgets, and cost per successful task into hard requirements. Engineer for fast iteration with safe sandboxes and clear approval workflows.

A.3 Summary Table

Table A.1: Production blockers and mitigations

Blocker	What it looks like	Engineering moves that work
Wrong problem selection	Agent built where a workflow or rules engine wins	Triage table, explicit "done", bounded action space
Data unreadiness	Missing IDs, stale KB, permission mismatch, poor retrieval provenance	Data contracts, entity resolution, retrieval test suites
Integration complexity	Auth failures, non-idempotent tools, partial side effects, brittle glue code	Tools as contracts, centralized tool gateways, policy gates outside agent
Wrong architecture	Long trajectories, multi-agent sprawl, no termination	Short runs, checkpoints, verification, single-agent first
Eval and observability gaps	No replay, no regression, cannot debug incidents	Harnesses, traces, online eval, shadow mode
Governance and security gaps	Prompt-only policies, over-privileged agents, no audit	Least privilege, approvals, deterministic blocks, audit trails
Org and adoption bottlenecks	Productivity gains do not compound, employee fear, low training	Redesign workflows, incentives, training, learning loops
Economics and procurement drag	Costs spike, latency hurts UX, reviews slow iteration	Budgets, caching, staged rollout, sandboxing, clear ownership

A.4 Failure Flow Diagram

Figure A.1: Where agent programs fail on the way to production

flowchart TB
  classDef ok stroke:#1f7a1f,stroke-width:1px,rx:10,ry:10;
  classDef risk stroke:#b45309,stroke-width:1px,rx:10,ry:10;
  classDef fail stroke:#b91c1c,stroke-width:1px,rx:10,ry:10;

  A["Use case selection<br/>Is an agent needed?"]:::risk
  B["Data readiness<br/>Can we retrieve the right facts with provenance?"]:::risk
  C["Integration readiness<br/>Tools, auth, idempotency, failure modes"]:::risk
  D["Architecture choice<br/>Short runs, checkpoints, verification"]:::risk
  E["Evaluation + observability<br/>Harness, replay, online monitoring"]:::risk
  F["Governance + security<br/>Least privilege, approvals, audit"]:::risk
  G["Operating model<br/>Ownership, training, change management"]:::risk
  H["Scaled production<br/>Measured outcomes + continuous improvement"]:::ok

  A -->|Wrong problem| X1["Pilot stagnates<br/>No measurable impact"]:::fail
  B -->|Bad context| X2["Hallucination or wrong action<br/>Trust collapses"]:::fail
  C -->|Brittle tools| X3["Incidents: partial writes<br/>or auth failures"]:::fail
  D -->|Unbounded loop| X4["Cost/latency runaway<br/>No termination"]:::fail
  E -->|No evidence| X5["Cannot debug regressions<br/>No safe rollout"]:::fail
  F -->|Weak controls| X6["Security/compliance block<br/>or real breach risk"]:::fail
  G -->|No redesign| X7["Local gains only<br/>No org-level lift"]:::fail

  A --> B --> C --> D --> E --> F --> G --> H

A.5 Agent Design Checklist

Problem Fit

[ ] Task requires adaptive multi-step behavior
[ ] Success criteria are measurable
[ ] Non-goals explicitly defined
[ ] Deterministic alternatives ruled out with rationale

Architecture

[ ] Scaffold type selected with documented rationale
[ ] Steps have clear entry/exit conditions
[ ] Tool allowlist defined per step
[ ] Stopping conditions are explicit
[ ] Checkpoint states exist for approvals
[ ] Recovery paths designed for failures

Tool Contracts

[ ] Each tool has typed input/output schema
[ ] Side effects documented (none/read/write)
[ ] Idempotency keys for write operations
[ ] Error responses defined
[ ] Auth scopes are least-privilege

Context Design

[ ] Context budget per step defined
[ ] Retrieval strategy documented
[ ] Memory boundaries clear (session/case/profile)
[ ] Sensitive data handling specified

Evaluation Readiness

[ ] Primary metrics defined
[ ] Non-negotiable thresholds set
[ ] Golden set structure planned
[ ] Judge approach selected

Governance Readiness

[ ] Risk tier assessed
[ ] Owner identified
[ ] Approval workflow known

A.6 Production Launch Checklist

Evaluation Gates

[ ] Golden set pass rate ≥ threshold: ____%
[ ] Zero non-negotiable violations
[ ] Regression suite passes
[ ] Adversarial/red-team suite passes
[ ] Edge case coverage verified

Operational Readiness

[ ] Observability configured
[ ] Traces captured
[ ] Metrics dashboards live
[ ] Logs aggregated
[ ] Alerts configured for:
[ ] Error rate spike
[ ] Latency degradation
[ ] Cost anomaly
[ ] Safety violations
[ ] Kill switch tested
[ ] Rollback procedure tested

Governance Gates

[ ] AI dossier complete
[ ] Evidence bundle assembled
[ ] Technical lead approved
[ ] Business owner approved
[ ] Risk/compliance approved (if required)
[ ] Security review complete (if required)

Rollout Plan

[ ] Rollout strategy defined (canary/phased/full)
[ ] Success criteria for each phase
[ ] Rollback triggers defined
[ ] Communication plan ready
[ ] Support/escalation contacts identified

Final Checks

[ ] Documentation current
[ ] Runbook available
[ ] On-call rotation set
[ ] Feedback collection enabled

A.7 Incident Response Checklist

Immediate (0-15 min)

[ ] Assess severity: ☐ P1 ☐ P2 ☐ P3
[ ] Decide: Investigate vs Mitigate first
[ ] If P1/P2: Activate kill switch or downgrade autonomy
[ ] Notify stakeholders per severity matrix
[ ] Preserve evidence (don't delete logs/traces)

Investigation (15-60 min)

[ ] Identify affected scope (users, time range, actions)
[ ] Pull relevant traces
[ ] Identify the specific run(s) that failed
[ ] Determine: Model issue vs Tool issue vs Data issue vs Policy issue

Root Cause

[ ] What was the input/context?
[ ] What did the agent decide?
[ ] What should it have decided?
[ ] Why did it decide wrong?
[ ] Retrieval failure
[ ] Prompt gap
[ ] Tool error
[ ] Policy not covered
[ ] Edge case not in training

Remediation

[ ] Immediate fix (config change, autonomy downgrade)
[ ] Add failing case to regression suite
[ ] Fix root cause (prompt, retrieval, tool, policy)
[ ] Verify fix passes regression
[ ] Staged rollout of fix

Post-Incident

[ ] Incident report completed
[ ] Lessons learned documented
[ ] Evaluation gaps identified
[ ] Process improvements identified
[ ] Stakeholder communication closed

A.8 Evaluation Planning Checklist

Metrics

[ ] Task success metric defined
[ ] Quality dimensions listed (accuracy, relevance, safety, style)
[ ] Non-negotiable thresholds set
[ ] Tolerance thresholds set

Datasets

[ ] Golden set size determined
[ ] Coverage matrix defined (by type, complexity, cohort)
[ ] Edge cases identified
[ ] Adversarial cases included
[ ] Production sample collection planned

Judges

[ ] Deterministic graders for verifiable outputs
[ ] LLM-as-judge for subjective quality
[ ] Human review protocol for calibration
[ ] SME involvement scheduled

Harness

[ ] Offline harness configured
[ ] Online harness configured
[ ] A/B framework available
[ ] Replay capability tested

Gates

[ ] Ship thresholds documented
[ ] Rollback triggers defined
[ ] Escalation path clear

A.9 Governance One-Page Template

Use this template for lightweight governance of medium-risk agents. For high-risk agents, expand to full AI Dossier per Chapter 12.

Agent Identity

Field	Value
Agent name
Version
Owner
Risk tier	☐ Low ☐ Medium ☐ High
Lifecycle state	☐ Pilot ☐ Limited Prod ☐ Full Prod

What It Does

Problem solved: (1-2 sentences)

Users:

Key actions: (list 3-5)

What it cannot do: (explicit boundaries)

Risk Summary

Risk	Mitigation
Data: What sensitive data?
Errors: What if it's wrong?
Misuse: How could it be abused?

Evaluation Evidence

Metric	Threshold	Actual	Pass?
Primary accuracy			☐
Safety violations	0		☐

Golden set size: _ cases Last eval date: ___

Controls

Control	Status
Autonomy level	☐ Advisory ☐ Approve ☐ Auto
Kill switch	☐ Configured ☐ Tested
Audit logging	☐ Enabled
Rollback plan	☐ Documented

Approvals

Role	Name	Date	Decision
Technical lead			☐ Approved
Business owner			☐ Approved
Risk/Compliance			☐ Approved

Review Schedule

Next review: _ Trigger for early review: ___

A.10 Cost Engineering Checklist

Anatomy

[ ] Cost per run broken down by component
[ ] Input vs output token ratio measured
[ ] Tool call costs quantified
[ ] Retrieval costs quantified

Optimization

[ ] Model routing by task complexity configured
[ ] Caching strategy for repeated queries
[ ] Context budgets per step enforced
[ ] Early termination for simple queries

Monitoring

[ ] Cost per successful task tracked
[ ] Cost anomaly alerts configured
[ ] Budget caps enforced
[ ] Runaway detection active

A.11 Case Study Template

Use this structure for documenting agent implementations.

Overview

Problem statement
Users and stakeholders
Systems integrated
Risk tier

Architecture

Scaffold type and rationale
Component stack (models, tools, memory)
Skill library
Key design decisions

Evaluation

Metrics and thresholds
Dataset composition
Judge approach

Governance

Risk tier
Evidence requirements
Lifecycle state

Outcomes

Results achieved
Lessons learned
Future improvements

A.12 Failure Prevention Checklist

Termination and bounds

[ ] Every loop has explicit termination criteria
[ ] Step limits are set and enforced
[ ] Cost budgets are defined per task
[ ] Failure states are designed, not just success paths

Verification and confidence

[ ] Outputs are verified before consequential actions
[ ] Confidence is calibrated to evidence quality
[ ] Uncertainty is surfaced in UX, not hidden
[ ] Citations are verified, not just attributed

Workflow and adoption

[ ] User research conducted with real workflows
[ ] Latency meets workflow requirements
[ ] Integration minimizes context switching
[ ] Success metrics align with user incentives

Feedback and improvement

[ ] Correction capture is designed into UX
[ ] Feedback flows to evaluation datasets
[ ] Error patterns are systematically reviewed
[ ] Improvements are validated against feedback

Navigation: Appendix Index | Appendix B →