## Sources

1. [The Case Against Building Your Own Agent Platform](https://www.oreilly.com/radar/the-case-against-building-your-own-agent-platform/)

---

### **The Case Against Building Your Own Agent Platform** by Pete Johnson

**Main Arguments**
*   **Systemic Underestimation of Scope:** Organizations frequently underestimate the complexity of building an internal AI agent platform, often conflating simple **workflows** with true **autonomous agents** [1-3]. 
*   **The Rapid Inversion of Build vs. Buy:** Historically, categories like container orchestration saw a slow shift toward third-party solutions, but the generative AI market inverted in just one year; internal builds dropped from 47% in 2024 to 24% by late 2025 [1, 4].
*   **Workflows vs. Agents:** There is a critical distinction between **workflows**, which use predefined code paths for LLMs, and **agents**, which dynamically direct their own processes and tool usage [5]. While building a workflow system internally may be tractable, the jump to supporting agents is not incremental and involves significant architectural hurdles [2, 3].
*   **The Four Pillars Trap:** Building an agent platform requires mastering four distinct, rapidly evolving product categories: **Memory, Governance, Evaluation (Eval), and Orchestration** [2, 6-9].
*   **Building the Wrong Thing:** The author argues that companies should build **business-specific agents** while buying the **technology-specific platform components** that support them [10].

**Key Takeaways**
*   **The "Long Tail" of Development:** Most internal projects fail because they do not account for the maturity curves and specialized expertise required for features like **temporal reasoning** in memory or **trajectory-based analysis** in evaluation [2, 6, 11].
*   **Legal and Security Risks:** The **EU AI Act**, becoming fully enforceable in August 2026, imposes strict requirements for audit trails and human oversight that many internal platforms are unprepared to meet [12]. Furthermore, traditional security like RBAC is insufficient for agents, which can be vulnerable to "excessive agency" and **indirect prompt injection** [7, 13].
*   **The Cost of Maintenance:** Internal platforms often bake in 2024 patterns (like flat vector retrieval) that become obsolete by 2026, forcing platform teams into constant **refactoring sprints** rather than shipping new features [14].
*   **Strategic Agility:** Enterprises that chose **model-agnostic platforms** in 2025 were able to adapt to massive shifts in market share between vendors like OpenAI, Anthropic, and Google, while those with hardcoded internal platforms were forced into costly rewrites [15, 16].

**Important Details**
*   **Memory Complexity:** Modern agent memory is not just a database; it requires three separate systems—**episodic, semantic, and procedural**—to handle temporal reasoning and fact validity [6].
*   **Evaluation Metrics:** Evaluation has shifted from simple accuracy to production metrics like **trajectory_exact_match** and **trajectory_precision**, which track the full path an agent takes to reach a result [8, 11].
*   **Orchestration Fragmentation:** There is no industry convergence on orchestration; frameworks like **LangGraph, CrewAI, AutoGen, and the Model Context Protocol** represent different, non-interchangeable bets on how agents should coordinate [9, 17].
*   **The Build Heuristic:** Building internally is only justified if an organization has a **durable competitive moat** through proprietary data (e.g., Mastercard or Plaid) or faces extreme regulatory hurdles that off-the-shelf tools cannot satisfy [18, 19].
*   **Personnel Risk:** Internal platforms create a "COBOL-like" legacy risk where the original "clever" team leaves, leaving the company to pay premium rates for contractors to maintain a custom, outdated system [20].
*   **Future Outlook:** Gartner predicts that over **40% of agentic AI projects will be canceled by 2027**, primarily due to scope estimation errors inherent in internal builds [21].