How Clay’s agent writes code its own engineers wouldn’t
Summary
Clay's new global agent emphasizes harness engineering over post-training, demonstrating that complex, outcome-oriented tasks can be achieved by building around a robust CLI and agent-friendly primitives. The system utilizes advanced techniques like isolated sandboxing (Firecracker/Daytona), high-durability orchestration (BullMQ), and a custom Knowledge Hub for memory, resulting in a highly expressive, though complex, agentic architecture.
Key takeaways
-
Shift from Task-Oriented to Outcome-Oriented Agents
12:12
The new Clay agent is designed to be comprehensive, handling end-to-end go-to-market tasks, contrasting with the previous, more granular, task-focused product, Sculptor (01:11).
-
Harness Engineering Outperforms Post-Training
The company found that the most significant performance gains came from squeezing juice out of the harness engineering and primitives, rather than achieving step-function differences through model post-training (00:00).
-
Complex Bash Commands and Security
The agent executes extremely complex, multi-line Bash commands (1:02:37). To ensure safety, the agent operates against a Firecracker isolated sandbox, with the agent running outside the sandbox (1:02:37).
-
Memory Precedes the Agent
The team prioritized building the memory system (Knowledge Hub) before the agent itself, recognizing the critical need to share business insights across the go-to-market motion (41:16).
-
Pricing Strategy: Tokens Over Outcomes
The company prefers charging for tokens (the unit of work) rather than outcomes, viewing tokens as a proxy for the work delivered, which is considered a more rational business model (56:57).
More digests on Agent Engineering
Technical details
-
Agent Architecture and Primitives
732s
The agent is built around a coding agent that leverages a custom CLI (Command Line Interface) as a core primitive. This allows the agent to perform complex operations like generating artifacts (workflows, audience segments) and scheduling follow-ups (07:32).
-
Orchestration and Durability
The agent relies on a robust orchestration layer, utilizing BullMQ to ensure high durability, allowing long-running tasks (30+ minutes) to continue even if a host fails (1:02:37).
-
Search and Data Layer
2153s
The system includes a custom SQL-like search query language (DSL) to construct complex queries against large datasets (900 million contacts, 40 million companies), addressing the limitations of simple filtering (21:53).
-
Memory and Context
Memory is managed through a two-tiered system: the company-level Knowledge Hub (a wiki-style understanding of the business) and agent-level memory (user-scoped, per thread). The Knowledge Hub is both auto-populated and self-updating (41:16).
-
Model Routing and Control
The system uses a post-trained model router, which resides at the harness layer. This allows the company to deliver excess value by matching user expectations to specific models, often achieving cost savings (59:14).
Mentioned resources
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.