IThe agent cemetery problem and three types of debt.
IICloud and infrastructure foundation.
IIIPlatform engineering as governance layer.
IVContext engineering and intent engineering.
VHarness engineering: the operating model.
VIToken economy, FinOps, and Azure AI Foundry.
Part I
I
The cemetery.
Why pilots die in production, and what is actually missing.
Problem · The gap
75% of enterprises plan agentic AI. Only34%report deep transformation.
+39%cognitive complexity from ungoverned agents (CMU AIDev)
Pilot agents fail to reach production for one consistent reason: they lack a coherent multi-layer engineering foundation. The cemetery grows because layers get skipped.
Problem · Three debts
Technical debt is familiar. Two new forms make it worse.
01 · Technical debt
Familiar accumulation.
Patches, fragile dependencies, brittle tests. Liu et al. analyzed 304,362 AI-generated commits across 6,275 repos and found higher shares of requirement and test debt vs human-written code.
02 · Cognitive debt
Cognitive surrender.
Developers accept AI-generated code without understanding it. Storey calls it cognitive surrender. The codebase becomes orphaned knowledge: nobody can modify it safely because nobody understands why it works.
03 · Intent debt
Objectives never captured.
The agent was built for metric X under constraint Y with hidden assumption Z. When production conditions shift, the agent optimizes for the wrong thing. Not broken, just unspecified.
Solution · Six disciplines
A stack that survives production.
Part II
II
Cloud foundation.
Kubernetes, MCP, autonomous infrastructure, the CPE 4-plane reference.
Cloud · Kubernetes
Kubernetes is the standard AI platform.
Why K8s wins for AI
GPU scheduling, autoscaling, multi-tenancy.
CNCF survey shows Kubernetes hosting AI workloads grew from niche to default. GPU operators, KServe, KubeRay, vLLM operators, all converge on K8s as the substrate.
What changes for agents
Long-running, stateful, bursty.
Agentic workloads are not stateless web requests. Sessions live for hours, hold KV cache, fan out to subagents. Pod design, network policies, and autoscaling need redesign.
Cloud · Reference architecture
Cognitive Platform Engineering: four planes.
Cloud · Protocol layer
MCP is the infrastructure protocol layer.
Why a protocol
N-by-M becomes N-plus-M.
Without a standard, every agent integrates every tool bilaterally. MCP collapses N agents times M tools into N plus M integrations.
What it carries
Tools, resources, prompts.
Three primitives. Tools execute side effects, resources expose data, prompts shape interaction. Identity and audit thread through every call.
Where it runs
Gateway pattern.
A platform-owned MCP gateway centralizes auth, scope, telemetry. Servers behind it carry policy, rate limits, and content classification.
Part III
III
Platform engineering.
Golden paths, guardrails, MCP server registries, agent RBAC.
Platform · Golden paths
The path of least resistance is the safe path.
Without golden paths
Every team rebuilds the same controls.
Each agent ships with its own auth, telemetry, secret handling. Inconsistent quality, audit holes, and governance becomes a paper exercise.
With golden paths
Compliant by default, custom by exception.
Backstage templates, IDP scaffolds, MCP server registries. Security and FinOps controls baked in. Teams opt out only when they need to.
Platform · Identity
Agents need identity, not service accounts.
SPIFFE
Workload identity.
Every agent gets a SPIFFE ID. Identity propagates from caller to MCP server to data source.
Scoped tokens
Least privilege per task.
Tokens minted per session, narrow scope, short TTL. The agent cannot exceed what the user could do.
Audit log
Independent of platform.
The auditor queries without engineering help. Decisions, identities, payloads classified, all traceable.
Policy
As code.
OPA, Cedar. Policies live in repos, version with the code, get peer review like infrastructure.
Part IV
IV
Context and intent.
What the agent knows, and what the agent should do.
Context vs Intent
Context alone is insufficient.
Context engineering
Smallest set of high-signal tokens.
Memory tiers, skills, MCP, AGENTS.md, RAG. Concave utility curve: more tokens degrade output past the peak. Compaction recovers 22.7%.
Intent engineering
Goals, values, trade-offs.
CONSTITUTION.md, Spec-Driven Development, written hierarchies of trade-offs. Without it, a well-contextualized agent optimizes for the wrong outcome.
Context · Concave utility
More tokens is not more value.
Context · Techniques
Six techniques that compound.
Compaction
22.7% reduction.
Structured compression of session history by the model itself. SWE-bench Lite paper, arXiv 2601.07190.
Tool result clearing
Stop paying twice.
Old tool results stay in context unless explicitly cleared. Recurring payment for content the agent already used.
E-mem subagents
-70% tokens.
Hierarchical memory across subagents. Parent keeps summaries, children carry detail. arXiv 2601.21714.
Persistent memory
11x compression.
Long-lived memory store outside the window. Retrieval brings only the relevant slice. arXiv 2603.13017.
SkillReducer
39% body, 48% description.
Programmatic reduction of skill files. Same behavior, smaller footprint. arXiv 2603.29919.
Prompt caching
Up to 90% input savings.
Reuse of KV cache from a stable prefix between requests. The largest single lever in most workloads.
Intent · Specification-driven
Spec-Kit makes intent legible.
CONSTITUTION.md
Non-negotiables.
Values and trade-off hierarchies the agent must respect across sessions. Not advice. Constraints.
SPECIFICATION.md
What to build.
Goals, success criteria, anti-goals. Reviewed before code is written. The agent reads it as input, not interpretation.
IMPLEMENTATION_PLAN.md
How.
The plan is artifact, not transient. Reviewable, version-controlled, replayable. Agents do not freelance the plan.
Part V
V
Harness engineering.
Where the four layers become a production system.
Harness · Why
A framework is not an operating model.
The four layers, infrastructure, platform, context, intent, describe what to build. The harness wraps them in cadence, ownership, evidence, and review. Without it, the framework lives in slides; with it, it ships.
Harness · Two paths
Three Horizons or Open Horizons.
Path A · Three Horizons Accelerator
Microsoft-aligned, opinionated.
Azure AI Foundry, GitHub Copilot Enterprise, Microsoft 365 Copilot, Teams. Pre-wired observability, FinOps, and identity. Faster start, narrower path.
Path B · Open Horizons
Open-source, multi-vendor.
CNCF stack, multi-model gateway, OpenTelemetry, FOCUS. More integration work, more sovereignty, more model flexibility. Same harness disciplines apply.
Integration · Maturity
Five maturity stages, exit criteria, real metrics.
Integration · Ten principles
Ten principles that bind the layers.
01Treat tokens as a finite resource. Budgets per session, agent, repo.
02Pre-compute compiler-grade context in CI, not per task.
03Identity propagates from caller to data source. No service accounts.
04Spec before code. CONSTITUTION before SPECIFICATION before plan.
05Audit logs are queryable without engineering help.
06Cache aggressively. Prefix, semantic, hierarchical, all three.
07Pick the smallest model that passes the eval, not the largest available.
08Workflow before agent. Agent only when adaptive control adds value.
09Observability is for the auditor, not just the engineer.
FinOps for AI, Azure AI Foundry, the June 2026 transition.
Token · The fact
OnJune 1, 2026, GitHub Copilot stops charging per request.
Premium Request Units retire. GitHub AI Credits take over. 1 AI Credit equals US$ 0.01. Input, output, and cached tokens are accounted at the published API rate of each model. Cost stops being per user intent and becomes per work performed.
FinOps for AI · Three phases
Inform, Optimize, Operate.
Inform
Visibility first.
Showback per session, agent, repo, team. FOCUS export. Power BI or Grafana dashboards. The team that cannot see the bill cannot reduce it.
Optimize
The disciplines compound.
Prompt caching, model routing, lean MCP, semantic cache, plan/execute. Eight applied patterns deliver 40 to 70% reduction in agentic workloads.
Operate
Budgets, alerts, anomaly detection.
Chargeback to teams. Anomaly detection on session cost. Quarterly review of budget vs actual. The harness ritual.
Foundry · When to use
Foundry, Copilot, or direct API.
Azure AI Foundry
Enterprise governance.
Model catalog, prompt flow, content safety, AAD/RBAC, PTU vs PAYG, private networking. Where regulated workloads live.
GitHub Copilot
Developer-facing surface.
VS Code Copilot, Coding Agent, custom prompts. AI Credits billing post-June. Where dev productivity lives.
Direct API
Custom agents.
Anthropic, OpenAI, vLLM via gateway. For agents the org builds, not the agents the org buys.
Token · Eight applied patterns
Eight patterns. 40 to 70% reduction.
01
Dynamic routing
Route by complexity. Haiku for trivia, Sonnet for code, Opus for plans.
02
Hierarchical cache
Prefix at provider, semantic at gateway, exact at app. Three layers compound.
03
Subagent fan-out
Parent keeps summary, children carry detail. E-mem hierarchy.
04
Plan-then-execute
Cheap model plans, capable model executes only the plan.
05
Semantic cache
Embedding similarity, not hash. Reuse near-duplicate queries safely.
06
Tool composition
Compose primitives instead of growing the tool catalog.
07
Lean MCP
Progressive disclosure. Tool schemas are billable; trim them.
08
Custom agent SDD
Spec-Kit driven. The plan is artifact, not transient.
Token · Anti-patterns
Five anti-patterns to stop today.
01Always picking the largest model. Pay 6x for output a smaller model would deliver.
02Treating an agent as a workflow. Adaptive control where deterministic control suffices.