Building an AI agent in an afternoon is now within reach of almost anyone in the enterprise with a credit card. The tools are accessible, the deployments are easy. The hard part is delivering the intended results. Gartner predicts that more than 40% of agentic AI projects will be canceled by 2027, and the EU AI Act Article 14 requirements for human oversight for high-risk AI systems take effect on August 2, 2026. The deciding factor for whether agentic AI reaches production isn't the model, the framework, or the use case. It's the infrastructure beneath the agent: the part the people building agents have never had to think about.

Organizations are racing to deploy agentic AI to stay competitive, which means pressure-testing is often overlooked. Every agent project should be scrutinized by three executives asking three different sets of questions. The CISO asks whether we are exposed. The CFO asks whether we are overspending. The chief AI officer asks whether we are getting value. As a product leader focused on AI governance, I see this pattern across customer environments. Three architecture layers answer those three questions: identity, observability, and cost optimization.

Why AI pilots stall

An agent is not a faster chatbot. It chains dozens of steps, calls external tools, retains state across sessions, and triggers real-world actions. Most inherit the credentials of whoever deployed them. They operate at machine speed without context for the consequences of each step.

The mismatch is not a competence gap on the human side. It is a time-horizon gap. An engineer reasons about a database change over hours. An agent triggers a hundred of them before anyone reviews the first. Traditional audit logging captures request and response. That does not catch this pattern.

When something breaks, the cost is rarely the incident. It is the months of stalled deployment that follow. The risk committee freezes pilots. The productivity gains the program was supposed to deliver never materialize. Finance still gets the API bill. Three architecture layers decide whether a deployment survives that pattern. Each one is the answer to a question the people building agents never had to ask.

Layer 1: Identity for non-human actors

Start with identity. The default failure looks routine: a product manager with broad API access spawns an agent that inherits the full scope of those credentials and runs at machine speed across systems no one inventoried.

The scale is bigger than most teams realize. Industry IAM research puts non-human identities at more than 100 to 1 versus human accounts, with some 2026 surveys putting the ratio as high as 144 to 1. A May 2026 Identity Gap Report found two-thirds are unseen and unmanaged. Agents are moving from human identities with their owners' permissions to first-class principals. They are purpose-bound, cryptographically attested, and scoped to one task at a time. Google's Agent Identity, built on SPIFFE, is one early example. The production pattern has three properties. Credentials are issued per agent task. Token lifetime is measured in minutes to hours, not weeks. Scope is narrowed to the specific tools and data classes the task requires, and the credential revokes automatically on task completion.

If a single static credential is good for a week and 50 different tasks, you are not running agentic AI. You are running a service account with extra steps.

Layer 2: Observability that serves all three executives

Identity controls what an agent can do. Observability shows what it's actually doing. One instrumentation layer, three views.

First, the security view. Traditional logging captures request and response, which assumes one human action per logged event. An agent's unit of work is a chain. Pick a tool, call it, read the result, decide the next step. Twenty steps, some of them writing to production. Instrument every step as a durable audit object, independently queryable. Understand which tool was invoked, what data was accessed, what policy applied, and what the agent reasoned to justify the next step. That's what Article 14 oversight requires for production.

Second, the business-outcomes view. Audit objects answer the CISO. The chief AI officer asks a different question. Is the agent accomplishing what we deployed it for, or burning compute on a tangent? An agent can run 200 tool calls, generate clean audit logs, and produce nothing. It might be looping on a sub-goal that drifted three steps back. Observe each step against the declared business purpose: on-task ratio, sub-goal coherence, progress markers. Project management telemetry for a non-human worker.

Third, the cost view. The same per-step instrumentation produces cost telemetry: token count per step, model per call, context size per turn, downstream tool-call costs. Without that attribution, the next section's optimizations are blind.

A busy agent and a productive agent look identical in the security log. They look identical on the bill too. The difference shows up only when all three views run from the same instrumentation.

Layer 3: Cost optimization

Cost is where the architecture pays back. Gartner's March 2026 analysis put agentic workloads at five to 30 times the token cost per task of a standard chatbot. The FinOps Foundation's 2026 State of FinOps report found that 73% of organizations exceeded their original AI budget projections. Three failure modes drive that overrun.

First, using the wrong model. Agents default to the most capable one available. They call a frontier model for tasks a smaller one could handle with identical quality: summarizing a transcript, formatting JSON, classifying a ticket. The RouteLLM paper at ICLR 2025 demonstrated that intelligent routing cuts total LLM inference cost 40% to 80% with no measurable quality loss on routine work. Move model selection from a per-developer choice to a per-policy layer.

Second, running in loops. Agents can spend without limit if no one is watching. A widely-cited 2026 incident saw a LangChain multi-agent system run an infinite loop for 11 days and burn $47,000 in API charges. Per-session token ceilings, loop-detection circuit breakers that flag tool calls highly similar to prior calls, and hard daily caps stop this before it generates the bill. In our deployments, a three-tier cost structure catches the bulk of runaway patterns: a $50 daily soft alert, a $100 daily hard cutoff forcing routing to cheaper models, and a $1,000 monthly ceiling requiring manager approval.

Third, re-paying for the same context on every step. Every step re-sends the accumulated system prompt and conversation history. By step 20 the agent has paid for that context 20 times. Vantage's 2026 analysis of agentic coding sessions found re-sent context accounts for roughly 62% of the average agent's bill, the biggest single optimization target in agentic workloads. Three patterns help: anchored summarization at phase boundaries, sliding context windows, and provider-native prompt caching at the gateway. Most agents skip caching entirely, though Anthropic prices cached input at roughly 10% of base, Gemini at 10% to 25%, and OpenAI at 50%.

Governing agent cost means seeing every call, every model, every token attributed to the agent and the business purpose. Then act on it. Token counts without business attribution tell you how many gallons of gas you burned, not where you drove.

The deployment velocity payoff

The three layers serve the three executive questions. Identity gates what the agent can do. Observability shows what it is doing. Cost optimization controls what it spends.

The honest counterargument is that governance always slows deployment. That is true when governance is bolted on as approval gates layered over an agent that wasn't built with observability or per-task identity. It is false when governance is built into the architecture from day one. Teams that experience governance as a brake installed the brake without the steering wheel.

Governance built right still costs something. Per-task credentials add work on every tool call. Observability infrastructure adds compute. The question is whether that cost beats the alternative.

The layers compound. Identity without observability is theoretical. Observability without cost control is descriptive. Without identity at the bottom, cost control becomes caps without context, forever reactive. All three together produce a governance review that runs in weeks, not quarters, because the data each executive needs already exists. In our experience, organizations with that infrastructure can deploy six workflows to production in the time competitors complete one governance review. The real ROI of agentic AI is not how much faster a single workflow runs. In practice, it's how many workflows your team can defensibly put into production in a year.

Before the next pilot

Here are four questions to run against any agent your team is about to push to production: Identity: For each agent in production, can you point to the per-task credentials it uses today, and the maximum scope of any single token? Observability: For any agent session, can you produce three views from the same instrumentation: the audit object per step, the on-task ratio versus tangents, and the per-step cost broken down by model and context size? Cost optimization: Does your platform automatically route by model, cap runaway loops, and avoid re-sending the same context every step? Velocity: How long does it take a new agent workflow to move from approved pilot to production in your environment today?

If the answer is months, the architecture above is the gap. Gartner's 40% stat is about your next pilot.

Source: InfoWorld News