Investors should shift focus from AI "brains" to "harness engineering" companies like Anthropic and Llama Index, which build the infrastructure and environments that allow AI models to execute complex, autonomous tasks. High-conviction opportunities exist in the software development space, where platforms like Cursor and Blitzy are achieving 5x engineering velocity by wrapping specialized harnesses around foundation models. Look for enterprise-grade "Managed Agents" from Anthropic to provide long-term stability for businesses, reducing the risk of custom code becoming obsolete as models evolve. In the fintech and compliance sectors, Mercury and Drata are top picks as they provide the modern API-driven backends and automated governance required for AI agents to operate safely. Expect a massive shift in enterprise operating models through 2026, favoring firms like KPMG that are aggressively redesigning their internal workflows to integrate human-AI collaboration.

Harness Engineering (The "General Harness")

Harness engineering is identified as the successor to prompt and context engineering. It refers to the systems, tooling, and infrastructure built around an AI model (the "brain") to give it "hands"—the ability to execute code, access specific data, and perform multi-step tasks autonomously.

• The Shift: In 2023/24, the focus was Prompt Engineering (how to talk to the model). In 2025, it shifted to Context Engineering (what data the model sees). Now, the focus is Harness Engineering (the environment the model operates in).
• Core Components: A standard harness consists of three layers:
  • Information Layer: Manages memory and tools (e.g., MCP servers, skills).
  • Execution Layer: Orchestrates how agents collaborate and provides "sandboxes" for safe code execution.
  • Feedback Layer: Handles evaluation, error recovery, and observability.
• The "Big Model" vs. "Big Harness" Debate: There is a tension between those who believe models will eventually become smart enough to not need complex scaffolding (OpenAI's Noam Brown) and those who believe specialized harnesses will always be necessary to solve complex business workflows (Llama Index).

Takeaways

• Investment Theme: Look for companies building "infrastructure for agents" rather than just the models themselves. The value is shifting from the "brain" to the "environment" that makes the brain useful.
• Efficiency Gains: Enterprises utilizing robust harnesses (like Blitzy) are reportedly seeing 5x engineering velocity, turning months of work into days.
• Product Convergence: Expect most software platforms (Notion, Linear, Microsoft) to look similar by 2026, as they all adopt a "general harness" architecture to turn their specific data into actionable agentic outcomes.

Anthropic (Claude / Managed Agents)

Anthropic is moving toward "Managed Agents," a hosted service that decouples the AI model from the execution environment.

• Managed Agents: This service provides a stable interface even as the underlying models and harnesses evolve.
• Obsolescence Risk: Anthropic noted that specific harness "hacks" (like context resets) become "dead weight" as models improve. Their strategy is to build a "meta-harness" that is unopinionated and adaptable.
• Claude Code: Highlighted as a "massive breakthrough" in looping agent architecture that generalizes well to any computer-based task.

Takeaways

• Platform Stability: For developers and enterprises, Anthropic’s "Managed Agents" reduces the risk of building complex custom scaffolding that might become obsolete with the next model update.
• Competitive Edge: Anthropic is positioning itself not just as a model provider, but as an infrastructure provider for "Long Horizon" agents.

Cursor (Anysphere)

The AI-native code editor Cursor is cited as a prime example of harness engineering integrated into a successful consumer product.

• Cursor 3: Introduced a "unified workspace" for agents. It moves the user from "manually editing files" to "managing fleets of agents."
• Feature Set: Includes multi-repo layouts and seamless handoffs between local and cloud agents, which are classic "harness" features.

Takeaways

• UX as a Moat: Cursor’s success suggests that the "Harness" (the user interface and agent orchestration) is just as important for user retention as the underlying model (GPT-4 or Claude).

Blitzy

An autonomous software development platform that focuses on "inference-time compute" and agent orchestration.

• Performance: Blitzy reported a 66.5% score on the Sweebench Pro benchmark, significantly outperforming GPT-4o (57.7%).
• The Thesis: Blitzy argues that a specialized harness wrapped around a foundation model can unlock higher performance than the model could achieve on its own. Their "knowledge graph" provides deep code-base context that raw models lack.

Takeaways

• Specialized vs. General: Blitzy serves as a case study for why specialized "vertical" harnesses (specifically for enterprise coding) may continue to outperform general-purpose models for high-stakes tasks.

Key Investment Sectors & Themes

1. Software Development Life Cycle (SDLC)

The podcast highlights a "third era of software development" where agents ship improvements autonomously.

• Opportunities: Companies like Cursor, Blitzy, and Cognition that are building the "harnesses" for this new era.

2. GRC (Governance, Risk, and Compliance)

• Mentioned Asset: Drata.
• Context: Using agentic platforms to automate tedious tasks like security questionnaires and evidence collection.
• Insight: GRC is a prime candidate for "Harness Engineering" because it requires high accuracy, access to many data silos (context), and repeatable workflows.

3. Enterprise Operating Models

• Mentioned Asset: KPMG.
• Context: KPMG is acting as "client zero," embedding agents across their enterprise to shift their total operating model.
• Insight: The "winner" in the enterprise space won't just be the one with the best AI tool, but the one that redesigns their organizational "harness"—the environment where humans and AI collaborate.

4. Fintech & Modern Banking

• Mentioned Asset: Mercury.
• Context: Business banking designed for modern software founders, emphasizing clean interfaces and API-like permissions.
• Insight: As AI agents begin to handle more business operations, they will require "modern" banking backends (like Mercury) that allow for granular digital control and permissions.