Track 2 · AI production engineering and AgentOps

Move one valuable AI-agent workflow from prototype to controlled production.

Peak Consulting is testing an Azure-based productionization service for companies whose agent prototypes are blocked by security, data access, evaluation, deployment, cost, observability, or unclear operating ownership.

Explore the offer → See the Azure platform foundation

1 named workflow 10–15 days readiness baseline 8–10 weeks launch project Human approval for consequential actions

Scope and prices are working hypotheses requiring separate buyer evidence.

The operating moment

The demo works. The production system does not exist yet.

Agentic systems are not only models and prompts. They combine business workflow, data, tools, authorization, application behavior, evaluation, telemetry, cost, and human intervention.

Control gap

Actions outrun permissions

The prototype can call tools, but nobody has defined what it may read, change, approve, retry, or never do.

Evidence gap

Demos replace evaluation

New prompts, tools, models, and indexes are judged by a few conversations rather than representative tests and release thresholds.

Operating gap

No owner can explain failure

Tracing, privacy, cost, escalation, rollback, and disablement are unresolved when the workflow reaches real users.

Prototype blocked by security Untracked agent tools No evaluation gate Uncontrolled trace data Cost and latency surprise

Commercial path

Bound the workflow before building the platform

The opening project productionizes one funded use case with named owners and reversible or human-approved actions. Additional agents and tool domains are separate decisions.

Step 1 · 10–15 business days

Agent Production Readiness Baseline

$15k–$25k working range

Assess the workflow, Foundry architecture, models, data, tools, identity, evaluation, tracing, privacy, cost, human approvals, and incident controls.

Output: target architecture, risk register, evaluation plan, control boundaries, and roadmap.

Step 2 · 8–10 weeks

Agent Launch Platform Project

$35k–$65k one workflow

Implement client-owned Azure/Foundry environments, deployment, approved knowledge and tools, evaluations, tracing, cost controls, approval, rollback, and runbooks.

Acceptance: the bounded workflow passes release gates and operates with a tested kill switch.

Step 3 · Fixed bridge

90-day Agent Stabilization

$8k–$15k/mo working range

Review traces and feedback, improve behavior through versioned changes, manage cost/latency, refine approval thresholds, and transfer ownership.

Not unlimited agent development or autonomous ownership of the business process.

Reference shape

The agent is only one layer of the production system

Microsoft Foundry supplies models, agents, evaluations, and governance capabilities. The workload still needs business boundaries, application controls, Azure foundations, and an operating system.

Business workflow

Owner Users Outcome Allowed actions Prohibited actions

Agent application

Orchestration State Instructions Approval experience

AI capabilities

Foundry models Agent Service Knowledge Approved tools

Azure controls

Projects and RBAC Managed identity Private network Policy and secrets

AgentOps

CI/CD Evaluations Traces Metrics and cost Incidents Rollback

The agent may propose an action. An external authorization control decides whether that identity may execute it, and a human approves consequential actions.

— Initial autonomy boundary

Production gates

Ten questions must have evidence before launch

Outcome

Who owns the result, who uses it, and what improves?

Scope

What may the agent do, never do, and escalate?

Data

Which sources, identities, sensitivity, freshness, and retention apply?

Tools

Are permissions minimal, inputs validated, actions auditable, and retries safe?

Quality

Which representative cases and thresholds block a weak release?

Safety

How are injection, abuse, unsafe outputs, and prohibited actions handled?

Platform

Are environments, IaC, network, secrets, policy, and capacity repeatable?

Operations

Can owners trace, alert, intervene, roll back, and disable the workflow?

Economics

What are latency and model/tool cost per completed task?

Launch

Who signs off, supports the release, and reviews measured value?

Implementation runbook

Build the evidence and controls in the same delivery loop

Phase	Human decision	Engineering output	Acceptance evidence
Week 0–1	Workflow, owners, prohibited actions, and success cases	Charter, risk register, evaluation dataset v1	Business, technical, and security owners approve boundaries
Week 2	Foundry, project, identity, and network decisions	IaC foundation and isolated non-production environment	Environment deploys repeatably
Week 3	Model, grounding, and data access	Model and knowledge integration	Source access, freshness, and grounding tested
Week 4	Tool and action boundaries	Least-privilege tool integrations	Unauthorized and malformed actions fail safely
Week 5–6	Quality, safety, trace, and privacy thresholds	Automated evaluations, tracing, and dashboards	Weak release is blocked; trace access/retention reviewed
Week 7	Intervention and failure controls	Human approval, rollback, disablement, and runbooks	Consequential action requires approval; kill switch works
Week 8–10	Bounded release and value review	Production launch, tuning, cost controls, handoff	Owner can operate, inspect, and stop the workflow

Initial use-case policy

Start where actions are reviewable and reversible

Better starting points

Internal knowledge with source links
Support research and response drafting
Document intake and classification
Engineering and incident investigation assistance
Compliance-evidence collection
Recommendations requiring human confirmation

Avoid at launch

Autonomous production infrastructure changes
Unrestricted shell, database, or cloud access
Final financial, legal, employment, or healthcare decisions
Payments, deletion, or commitments without approval
Broad cross-system write identities
Open-ended multi-agent autonomy without clear ownership

Automation boundary

AI may help build the system. It cannot approve itself.

AI can accelerate

Architecture and IaC scaffolds
Evaluation-case and edge-case drafts
Trace, cost, and failure-pattern analysis
Prompt/tool documentation
Runbook and threat-scenario drafts

Humans must own

Business acceptance criteria
Trust boundaries and allowed actions
Production identities and permissions
Privacy, retention, and regulatory decisions
Release approval and incident response

Tracing is customer data.

Agent traces may contain prompts, responses, retrieved context, tool calls, intermediate steps, personal information, latency, tokens, and errors. Enable tracing deliberately; redact sensitive data; restrict access; and define retention before production.

Current references: Microsoft Foundry architecture, Foundry landing-zone baseline, tracing and data handling, and Agent Service transparency guidance.

AgentOps begins with the same Azure foundation as platform engineering.

Identity, network, IaC, CI/CD, observability, cost, recovery, and client ownership remain the common operating core.

Explore Azure Platform Engineering → See the validation method