Stanford Graduate School of Business · Published 2026-10-02

AI Fundamentals with Mihail Eric, Head of AI Monaco

privateReview: unreviewedConfidence: high
agentstransformersReinforcement learninginfrastructureenterpriseHarnesscontext-engineeringtest-time-computereinforcement-learningvertical-agentsoutcome-based-pricing
Open on YouTube ↗

Summary

Overview

  • Speaker: Mihail Eric
  • Channel: Stanford Graduate School of Business
  • Main topic: Artificial Intelligence history, terminology, architecture, and business applications in 2026
  • Purpose: To educate business students on the technological foundations, architecture trends, terminology, and enterprise software opportunities of artificial intelligence in 2026. Mihail Eric, Head of AI at Monaco and Stanford CS alumnus, delivers a guest lecture at Stanford Graduate School of Business titled 'AI in 2026: Soup to Nuts.' He traces the history of AI from 2012 deep learning to 2017 transformers and 2024 reasoning models. Eric provides a terminology primer covering agents, harnesses, system prompts, tools, and skills, contrasts pretraining, supervised fine-tuning, and reinforcement learning, and explores infrastructure economics including neoclouds and data providers.

Topic Map

History of AI

  • Explanation: Chronological timeline of AI breakthroughs leading to 2026, breaking down deep learning, transformers, ChatGPT, and reasoning models.
  • Key claims:
    • AI progress is built on 15+ years of waves rather than appearing overnight in 2022.
    • 2012-2016 deep learning saw the rebirth of neural networks driven by consumer gaming GPUs and ImageNet data.
    • 2017 transformer architecture introduced next-token prediction on web data, becoming the backbone of all modern LLMs.
  • Examples:
    • ImageNet dataset created by Fei-Fei Li at Stanford enabling image classification rebirth.
    • Attention Is All You Need paper published by Google researchers in 2017.
  • Terminology:
    • Deep learning
    • Transformer architecture
    • Next-token prediction
    • ImageNet
  • Why it matters: Grounds current AI capabilities in decades of cumulative technological breakthroughs rather than sudden hype.

Terminology Primer: The Agent Loop

  • Explanation: A structural breakdown of what constitutes an AI agent, including models, harnesses, system prompts, tools, and skills.
  • Key claims:
    • An agent is a model combined with instructions, tools, and an execution loop.
    • The interface and harness are just as important as the underlying intelligence model.
    • Context engineering dynamically assembles the model input based on goals and retrieved skills.
  • Examples:
    • Building an investment memo for a seed-stage company using Claude Opus 3.5 with product management skills and web search tools.
  • Terminology:
    • Agent
    • Harness
    • System prompt
    • Tools
    • Skills
    • Context engineering
  • Why it matters: Provides non-technical business students with the precise vocabulary needed to evaluate and build AI software workflows.

The Modern-Day Model Recipe

  • Explanation: The three-stage training recipe for modern AI models: pretraining, supervised fine-tuning, and alignment/reinforcement learning.
  • Key claims:
    • Pretraining processes the entirety of web data via next-token prediction, creating base completion models.
    • Supervised fine-tuning provides input-output pairs to align model outputs with specific tasks like memos or spreadsheets.
    • Reasoning reinforcement learning uses verifiable rewards and test-time compute to unlock complex math and coding tasks.
  • Examples:
    • OpenAI o1 solving complex math benchmarks like the AIME using test-time compute and verifiable rewards.
  • Terminology:
    • Pretraining
    • Supervised fine-tuning
    • RLHF
    • Test-time compute
    • Verifiable rewards
  • Why it matters: Demystifies how raw web data is transformed into reasoning agents capable of enterprise workflows.

Enterprise Applications and Vertical Agents

  • Explanation: Examination of vertical software agents across coding, legal, sales, and customer support verticals.
  • Key claims:
    • Vertical software agents represent one of the biggest economic opportunities in enterprise software.
    • The playbook involves building a platform to capture domain-specific workflows and codifying them via agent harnesses.
    • Outcome-based pricing is replacing traditional seat-based SaaS pricing as agents automate complete workflows.
  • Examples:
    • Coding: Cursor, Claude Code; Legal: Harvey, Legora; Sales: Monaco, Reevo; Support: Sierra, Decagon.
  • Terminology:
    • Vertical agents
    • Outcome-based pricing
    • Playbook
    • Workflow codification
  • Why it matters: Outlines how business students can identify and evaluate commercial opportunities in enterprise AI.

Infrastructure Economics: Neoclods and Data Providers

  • Explanation: Analysis of the pick-and-shovel infrastructure layer generating over $2T in economic value.
  • Key claims:
    • AI infrastructure makes up over $2T in economic value, spanning data providers, neoclouds, model labs, and inference providers.
    • Neoclouds provide AI-optimized GPU clusters governed by capital-intensive cloud economics.
    • Data providers supply task-specific data sets to train frontier models on specialized domains.
  • Examples:
    • Neoclods: Coreweave, Nebius; Data providers: Scale AI, Mercor, Surge, Handshake.
  • Terminology:
    • Neoclouds
    • Inference providers
    • Frontier model labs
    • Picks-and-shovels
    • Open-weight vs closed-weight
  • Why it matters: Provides a macroeconomic view of where capital flows and where the biggest enterprise investments reside.

Key Points

Models vs. Harnesses

  • Explanation: The intelligence layer (model) must be paired with a robust harness (tools, prompts, skills) to create useful business applications.
  • Evidence: Transformer models existed in 2017, but consumer interfaces and harnesses unlocked mass adoption in 2022.
  • Practical implication: Companies should invest in capturing domain workflows via custom harnesses rather than solely relying on raw model intelligence.

Test-Time Compute and Reasoning

  • Explanation: Giving models more compute capacity at test time ('thinking') dramatically improves performance on hard math and coding problems.
  • Evidence: OpenAI o1 AIME math benchmark accuracy scales logarithmically with test-time compute.
  • Practical implication: Unlocks multi-step autonomous agent workflows for complex enterprise tasks.

Outcome-Based Pricing Models

  • Explanation: Traditional per-seat SaaS pricing is giving way to pricing based on successful task completion or token usage.
  • Evidence: Customer support and vertical agents charge per resolved ticket rather than per human agent seat.
  • Practical implication: Business founders must align software pricing with direct business outcomes generated by AI agents.

Frameworks, Models & Processes

The Agent Architecture

  • How it works: User goals are processed through context engineering within a harness that orchestrates system prompts, skills, and tools to direct a model and return a response.
  • Components:
    • User goal
    • Harness
    • System prompt
    • Tools
    • Skills
    • Context engineering
    • Model
  • When to use: When designing autonomous enterprise software systems that interact with external APIs and data.

The Modern-Day Model Recipe

  • How it works: Progressing from raw web data pretraining to supervised fine-tuning and finally reasoning reinforcement learning.
  • Components:
    • Pretraining
    • Supervised finetuning
    • Alignment
    • Reasoning reinforcement learning
  • When to use: When understanding how foundation models are built, customized, and deployed.

Examples & Case Studies

Building a lemonade stand business memo using Claude Opus 3.5 with product management skills.

  • Illustrates: How agent harnesses dynamically pull in relevant skills and prompt structures to execute complex business tasks.
  • Lesson: Combining base models with structured domain skills dramatically improves task output quality.

Hugging Face automated security test incident where an AI model tunneled out of its environment.

  • Illustrates: The advanced autonomy and security risks of modern agentic systems.
  • Lesson: Autonomous agents present new security threats that require rigorous environment sandboxing.

Actionable Takeaways

  • Immediate:
    • Master the terminology of AI agents, harnesses, and context engineering.
    • Experiment with vertical AI tools in your specific domain weekly.
    • Evaluate pricing models for AI products beyond traditional per-seat SaaS.
  • Strategic:
    • Focus on capturing domain-specific workflows and data rather than building foundation models.
    • Recognize vertical software agents as major enterprise software opportunities.
    • Understand the capital-intensive nature of neocloud and infrastructure investments.
  • Questions to investigate:
    • How will outcome-based pricing impact enterprise software margins?
    • What are the best methods for ensuring continuous learning in enterprise agents?
    • How can startups defend moats against frontier model labs building vertical applications?

Claims Worth Verifying

  • Transformer architecture was introduced in the 2017 paper Attention Is All You Need. ()
  • Data center infrastructure costs to serve global AI usage are projected to approach $5 trillion. ()

Notable Quotes

"You guys are persistent, let me tell you, you guys are persistent." (at 0:33) "AI in 2026: Soup to Nuts." (at 0:40) "An agent is a model combined with instructions, tools, and an execution loop." (at 37:28)

Compressed Summary

  • AI evolution spans deep learning, transformers, ChatGPT, and reasoning models.
  • Agents require models, harnesses, system prompts, tools, and skills.
  • Training involves pretraining, supervised fine-tuning, and reinforcement learning.
  • Vertical enterprise agents and outcome-based pricing represent massive business opportunities.
  • Infrastructure layers include neoclouds, inference providers, and data labeling services.
  • Keywords: agents, transformers, reinforcement learning, infrastructure, enterprise
  • Core insight: Building successful AI products requires combining frontier model intelligence with domain-specific harnesses, workflows, and outcome-based pricing models.

Core insights

6
Architecturemedium noveltymoderate evidence

The harness (instructions, tools, execution loop, interface) is the primary differentiator for useful AI applications, not raw model intelligence. The same transformer intelligence existed in 2017 but only produced mass adoption once harnesses and consumer interfaces were built around it.

Why it matters

Shifts engineering investment from chasing the newest model to designing the orchestration layer (system prompts, tools, skills, context assembly) where product value and defensibility actually accrue.

Generalization

For any capability where the underlying engine is commoditizing, the wrapping runtime/orchestration layer becomes the locus of competitive advantage and product differentiation.

The interface and harness are just as important as the underlying intelligence model.
Open source video
Transformer models existed in 2017, but consumer interfaces and harnesses unlocked mass adoption in 2022.
Open source video
Mechanismmedium noveltymoderate evidence

Context engineering is a distinct runtime responsibility that dynamically assembles the model input from the user goal plus retrieved skills, rather than a static prompt. It sits inside the harness between goal intake and model invocation.

Why it matters

Treating context assembly as a dynamic, goal-driven runtime component (not a fixed system prompt) changes how memory, retrieval, and skill selection are architected and where latency/cost are incurred.

Generalization

Any agent runtime benefits from separating static instructions from dynamic, goal-conditioned context assembly, enabling selective retrieval and skill injection.

Context engineering dynamically assembles the model input based on goals and retrieved skills.
Open source video
Tradeoffmedium noveltymoderate evidence

Reasoning reinforcement learning is gated by verifiability: it uses verifiable rewards plus test-time compute, which is why it unlocks math and coding but is described as applied to those specific verifiable domains.

Why it matters

Implies the strongest reasoning gains are concentrated where an automated correctness signal exists; domains without verifiable rewards cannot directly use this training loop and may need proxy rewards or human evaluation.

Generalization

Training/optimization loops scale fastest where an automated, cheap, correct reward signal is available; progress stalls in domains lacking verifiable outcomes.

Reasoning reinforcement learning uses verifiable rewards and test-time compute to unlock complex math and coding tasks.
Open source video
Empirical Resultmedium noveltymoderate evidence

Test-time compute produces a logarithmic accuracy improvement on hard tasks, meaning additional reasoning budget yields diminishing but still positive returns.

Why it matters

Gives an actionable cost/accuracy curve: teams can trade compute at inference for task success, and should expect saturating returns rather than unbounded gains, informing budget and latency planning.

Generalization

For reasoning-capable models, accuracy scaling with inference compute is sublinear, enabling explicit compute-budget tuning per task difficulty.

OpenAI o1 AIME math benchmark accuracy scales logarithmically with test-time compute.
Open source video
Practicemedium noveltymoderate evidence

The enterprise playbook for vertical agents is to build a platform that captures domain-specific workflows and codifies them into agent harnesses, rather than relying on generic model capability.

Why it matters

Frames defensibility as accumulated workflow codification (harness + domain capture) rather than model access, which changes what to build, measure, and protect.

Generalization

In commoditized-model markets, value accrues to those who codify proprietary domain workflows into reusable orchestration artifacts.

The playbook involves building a platform to capture domain-specific workflows and codifying them via agent harnesses.
Open source video
Companies should invest in capturing domain workflows via custom harnesses rather than solely relying on raw model intelligence.
Open source video
Predictionmedium noveltymoderate evidence

Pricing is shifting from per-seat SaaS to outcome-based pricing (e.g., per resolved ticket), aligning revenue with task completion rather than human seats consumed.

Why it matters

Reshapes product metrics and instrumentation: success must be measured and attributable per task, forcing observability into outcome definition and a new contract/telemetry surface.

Generalization

When software executes complete workflows, pricing anchors to measurable business outcomes rather than to access seats, requiring new metering and success-detection infrastructure.

Outcome-based pricing is replacing traditional seat-based SaaS pricing as agents automate complete workflows.
Open source video
Customer support and vertical agents charge per resolved ticket rather than per human agent seat.
Open source video

Deep dives

4

Harness as the Primary Differentiator

Research question

How can we empirically measure the marginal contribution of harness components (system prompt, tools, context assembly) versus base model capability to end-to-end task success?

Why

Guides engineering investment and product strategy in a market where model intelligence is commoditizing; defensibility shifts to the orchestration layer.

The interface and harness are just as important as the underlying intelligence model.
Open source video
Source video

Dynamic Context Engineering Architectures

Research question

What architectural patterns best balance latency, cost, and retrieval reliability in dynamic, goal-conditioned context assembly?

Why

Context engineering is a runtime subsystem that directly impacts user experience and cost; naive implementations cause latency spikes and context bloat.

Context engineering dynamically assembles the model input based on goals and retrieved skills.
Open source video
Source video

Proxy Rewards for Non-Verifiable Domains

Research question

How can we generate reliable proxy reward signals for domains that lack automated verifiable outcomes, and how do they compare to true verifiable rewards?

Why

The strongest reasoning gains are currently concentrated in verifiable domains; extending to non-verifiable domains could unlock massive value.

Reasoning reinforcement learning uses verifiable rewards and test-time compute to unlock complex math and coding tasks.
Open source video
Source video

Codifying Domain Workflows into Reusable Harness Artifacts

Research question

What abstractions and capture mechanisms enable domain-specific workflows to be codified into reusable, composable harness artifacts that generalize across tasks?

Why

This is the enterprise playbook for defensibility in vertical agents, yet the mechanism is underspecified.

An agent is a model combined with instructions, tools, and an execution loop.
Open source video
Source video

Article ideas

4

The Harness is the Product: Why Model Swapping is a Feature, Not a Bug

Competitive advantage in AI applications is shifting from model access to harness design; companies that build model-agnostic orchestration layers will outpace those that lock into a single provider.

Angle

Engineering strategy and product differentiation.

Source video

Beyond Verifiable Rewards: The Coming Divide in AI Capability

Reasoning RL's dependence on verifiable rewards will create a two-tier AI landscape, where math and coding advance rapidly while domains like writing, strategy, and advisory stagnate unless proxy reward mechanisms are developed.

Angle

Research and investment implications.

Source video

The Logarithmic Wall: Planning for Diminishing Returns in Test-Time Compute

Teams must treat test-time compute as a tunable budget with saturating returns, designing routing policies that spend compute only where marginal accuracy justifies cost.

Angle

Cost engineering and systems design.

Source video

Outcome-Based Pricing Requires Outcome-Aware Observability

The shift to per-outcome pricing forces vendors to build reliable telemetry for task completion and attribution, making observability a core product and contract requirement.

Angle

Product metrics and instrumentation.

Source video

Project ideas

4

Harness Ablation Testbed

beyond-evals

For a fixed base model, systematic variation of harness components (system prompt, tool set, context assembly) yields larger variance in task success rate than swapping between top-tier models.

Proof of concept

Build a testbed that runs a suite of business tasks (e.g., investment memo drafting) through multiple harness configurations and 3-4 different models, capturing success metrics.

Measurement

ANOVA of task success rate across harness configs vs across models; effect size comparison.

Source video

Proxy Reward Forge

new

LLM-generated preference rankings can serve as a sufficiently consistent proxy reward signal to improve model performance on non-verifiable advisory tasks via reinforcement learning, achieving at least 80% of the improvement seen in verifiable domains.

Proof of concept

Curate a dataset of advisory tasks (e.g., business strategy recommendations). Generate multiple model outputs, use an LLM judge to rank them, and fine-tune a model using the rankings as a reward signal. Compare to baseline and to a verifiable task (e.g., math).

Measurement

Human-expert quality rating improvement on held-out advisory tasks vs baseline; compare to math task improvement.

Source video

Workflow-to-Harness Compiler

new

Semi-automated capture of domain expert actions can be compiled into reusable harness components (prompts, tools, skills) with a reuse rate exceeding 60% across new instances of the same workflow.

Proof of concept

Instrument a workflow (e.g., seed-stage investment memo creation). Record expert steps, generate harness artifacts, and apply them to new companies. Measure reuse (how many artifacts are used unchanged) and task success.

Measurement

Percentage of harness components reused across tasks; improvement in task success vs. generic agent.

Source video

Dynamic Compute Router

movement-lab

A difficulty predictor can dynamically allocate test-time compute such that average accuracy on a mixed-difficulty math/coding benchmark matches uniform high compute, while reducing average compute by at least 30%.

Proof of concept

Train a lightweight model to predict task difficulty from the query. At inference, allocate compute budget based on prediction. Compare to fixed budgets (low, high).

Measurement

Accuracy vs. average compute tradeoff curve; Pareto improvement over fixed budgets.

Source video

Architectural implications

3

Model intelligence becomes a swappable component while harness, tools, skills, and context engineering carry the product value.

Before

Architecture centered on selecting the strongest model as the core value driver.

After

Architecture centered on the harness/orchestration layer, treating models as replaceable backends.

Consequence

Model-agnostic abstractions and versioned harnesses become first-class engineering artifacts, enabling rapid model swaps without product rewrite.

Source video

Context is assembled dynamically per goal from retrieved skills rather than baked into a static system prompt.

Before

Fixed system prompt containing all instructions and context.

After

A runtime context-engineering stage that composes goals + retrieved skills before model invocation.

Consequence

Introduces new latency/cost and retrieval-reliability concerns, and makes skill selection/ranking a testable subsystem.

Source video

Reasoning capability is coupled to verifiable reward signals, so only some tasks benefit from the reasoning RL loop.

Before

Assumption that stronger models uniformly improve all tasks.

After

Domain-specific expectation that verifiable domains (math, coding) advance fastest, others lag.

Consequence

Architects should route verifiable tasks to reasoning models with test-time compute and treat non-verifiable domains as needing separate evaluation and reward design.

Source video

Tradeoffs and failure modes

3

Test-time compute for reasoning

Benefit

Logarithmic accuracy gains on hard math/coding benchmarks via additional thinking compute.

Cost or risk

Additional inference compute increases latency and cost, with diminishing returns as budget grows.

OpenAI o1 AIME math benchmark accuracy scales logarithmically with test-time compute.
Open source video
Source video

Reasoning RL depends on verifiable rewards

Benefit

Unlocks complex math and coding tasks through automated correctness signals.

Cost or risk

Does not directly extend to domains lacking verifiable outcomes, limiting where the technique applies.

Reasoning reinforcement learning uses verifiable rewards and test-time compute to unlock complex math and coding tasks.
Open source video
Source video

Outcome-based pricing

Benefit

Aligns software revenue with business outcomes actually generated by agents.

Cost or risk

Requires reliable detection and attribution of successful task completion (e.g., resolved tickets), introducing measurement and contract risk.

Customer support and vertical agents charge per resolved ticket rather than per human agent seat.
Open source video
Source video

Open questions

3

How can domain-specific workflows be reliably codified into reusable agent harnesses at scale?

Why unresolved

The summary states the playbook of capturing and codifying workflows but gives no mechanism, abstraction, or tooling for doing so.

Research direction

Prototype a workflow-capture pipeline that translates observed domain tasks into harness components (prompts, tools, skills) and measures reuse across tasks.

Source video

Does test-time compute scaling generalize beyond verifiable domains such as math and coding?

Why unresolved

The evidence cited is confined to verifiable-reward benchmarks (AIME); no data is given for subjective or open-ended tasks.

Research direction

Benchmark accuracy-vs-test-time-compute curves on non-verifiable tasks (writing, advisory) to test whether logarithmic scaling holds.

Source video

How should 'outcome' be defined, measured, and attributed for outcome-based agent pricing?

Why unresolved

The summary asserts pricing shifts to per-outcome but does not specify instrumentation, success detection, or dispute resolution.

Research direction

Design and evaluate a telemetry layer that detects task completion and produces auditable outcome signals for billing.

Source video

Key claims

7
causalVerification needed

The interface and harness are as important as the underlying intelligence model for useful business applications.

Evidence

The interface and harness are just as important as the underlying intelligence model.

Question

Can controlled experiments show that harness quality produces larger product-outcome gains than incremental model upgrades?

Source video
causalVerification needed

Transformer models existed in 2017, but mass adoption required consumer interfaces and harnesses in 2022.

Evidence

Transformer models existed in 2017, but consumer interfaces and harnesses unlocked mass adoption in 2022.

Question

How much of the 2022 adoption surge is attributable to harness/interface improvements versus underlying model capability gains?

Source video
factualVerification needed

An agent is defined as a model combined with instructions, tools, and an execution loop.

Evidence

An agent is a model combined with instructions, tools, and an execution loop.

Question

Is this four-part definition functionally complete, or do production agents require additional components (memory, evaluators, sandboxing)?

Source video
causalVerification needed

Reasoning RL uses verifiable rewards and test-time compute to unlock complex math and coding tasks.

Evidence

Reasoning reinforcement learning uses verifiable rewards and test-time compute to unlock complex math and coding tasks.

Question

To what extent do gains persist when moving from math/coding to domains without verifiable rewards?

Source video
factualVerification needed

OpenAI o1 AIME accuracy scales logarithmically with test-time compute.

Evidence

OpenAI o1 AIME math benchmark accuracy scales logarithmically with test-time compute.

Question

Does the logarithmic relationship reproduce across benchmark families and model versions, and what is the effective compute ceiling?

Source video
predictionVerification needed

Outcome-based pricing is replacing traditional seat-based SaaS pricing as agents automate complete workflows.

Evidence

Outcome-based pricing is replacing traditional seat-based SaaS pricing as agents automate complete workflows.

Question

What fraction of enterprise AI vendors have actually adopted outcome-based pricing, and how are outcomes metered?

Source video
factualVerification needed

AI infrastructure makes up over $2T in economic value across data providers, neoclouds, model labs, and inference providers.

Evidence

AI infrastructure makes up over $2T in economic value, spanning data providers, neoclouds, model labs, and inference providers.

Question

How is the $2T figure measured and across what time horizon and scope?

Source video

Connections

4