Rebuilding Omni: Transitioning to a Coding Agent for Airtable
Airtable evolves its AI agent from rigid workflows to a flexible coding-agent architecture.
1Limitations of Initial Design
The original Omni, released in summer 2025, functioned as a workflow system powered by LLMs. It relied on hard-wired subagents for specific tasks like OKR queries or interface modifications. This architecture struggled with open-ended scenarios, such as debugging complex automations, because the system could not dynamically adapt to multi-step, iterative investigation processes.
2The Shift to Agentic Models
By early 2026, the engineering team observed a seismic shift in coding agents capable of discerning ambiguous intent and self-correcting. To match this, they decided to rewrite Omni entirely. The new version abandons custom-built frameworks in favor of a battle-tested harness that manages agent loops, tool calling, and context management more effectively.
3Core Architectural Components
The new Omni architecture provides three essential elements: source code access, a working environment, and execution feedback. By utilizing Airtable MCP, the agent can inspect table schemas, interfaces, and automation scripts. An isolated sandbox with Python and Unix utilities allows the agent to execute code and analyze logs to determine subsequent actions.
4Efficiency via File-Based Tools
To handle large data imports, the team implemented file-based tools instead of generating arguments token by token. Omni now passes file paths to a local MCP server, allowing it to sample data rather than loading entire spreadsheets into context. This transition reduced the time for a 1,000-row import from one hour to minutes.
5Programmatic Tool Execution
Omni now leverages programmatic tool calling, enabling it to write Python scripts that process data directly. For tasks like spreadsheet imports or interface updates, the agent writes scripts to transform values or edit specific source lines. This approach avoids the high cost of inference-based generation and prevents context window exhaustion during complex operations.
6Encoding Expert Knowledge
Beyond basic functionality, the team introduced 'skills' to encode production-ready knowledge into Omni. These skills provide instructions for handling large imports and building mobile-friendly interfaces. By embedding these best practices, the agent moves closer to being a specialized expert capable of preparing applications for thousands of users, rather than just a generic assistant.
7Measurable Performance Gains
The rewrite yielded significant improvements in user engagement and satisfaction metrics. Weekly active users increased by 60% following the update. Furthermore, positive user ratings for the agent rose from 75% to over 90%. Feedback indicates that users are successfully leveraging the new Omni for complex automation fixes and custom interface development tasks.
8Future Extensibility Potential
The current architecture allows for continuous improvement as new models and tools are integrated. The team envisions a future where teams can add their own organizational skills, teaching Omni specific conventions and internal practices. This evolution aims to transform the agent from a generic tool into a highly personalized expert for every unique business.
9Strategic Significance of Agents
The transition to a coding-agent model represents a fundamental shift in how users interact with complex platforms like Airtable. By treating app configuration as code, the system gains the flexibility required for autonomous problem-solving. This approach validates the necessity of robust, tool-rich environments for AI agents to deliver reliable, production-grade results for users.