Back office ops · Production

How Airtable built its AI agents framework powering Omni and Field Agents

The problem

Airtable's original AI features were limited to simple generative use cases and could not reason through problems requiring dynamic decision-making, retrieve additional data beyond what was provided upfront, or support a conversational interface for user feedback and follow-up.

First attempt

The original AI capabilities — the AI field, AI in automations, AI-generated select options, and AI formula generation — were incapable of dynamic reasoning or retrieving additional data, limiting them to straightforward generative tasks.

Workflow diagram · grounded in source

User message triggers agent loop

trigger

“A user message event is produced by the user sending a message into the interaction. This triggers the decision engine to call the backing LLM to decide what step to take next.”

Context manager assembles world state

ai_action

“The context manager maintains all of the information accessible by the agent to accomplish the user's task. It's responsible for 'remembering' guidelines context (instructions for tone and purpose), session context (Airtable base, worksp…”

Decision engine calls LLM

ai_action

“The decision engine takes in all of the context from the context manager and decides the next step to take in the agent loop. This is the 'brain' of the agent and dictates the control flow via an LLM, human-in-the-loop, or a fixed workflow.”

Tool dispatcher executes tools

integration

“The tool dispatcher runs tools requested by the decision engine. When the tool call is finished running, the tool dispatcher sends a tool call output event to the agent. Depending on the agent's purpose, the tool dispatcher exposes a dif…”

LLM self-corrects on errors

feedback_loop

“the tool call failure information is passed back to the LLM. In most cases, with a descriptive error message, the LLM can self-correct and re-run the tool with different arguments, or provide the user with a reason as to why their reques…”

Final response ends agent loop

output

“An LLM message event is produced by the decision engine when the backing LLM produces a final output message. This ends the agent loop.”

Reported outcome

Airtable built a custom asynchronous event-driven state machine agentic framework that powers Omni and Field Agents, enabling reasoning, planning, and multi-step orchestration, with context optimization achieving a 15–30% reduction in tokens, inference latency, and cost.

Reported metrics

token, inference latency, and cost reduction from ID aliasing~15–30%

work automatable by AI agentsthousands of hours of work in seconds

Reported stack

OpenAIAnthropic

Source

https://medium.com/airtable-eng/how-we-built-ai-agents-at-airtable-70838d73cc43

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Airtable built a custom asynchronous event-driven state machine agentic framework that powers Omni and Field Agents, enabling reasoning, planning, and multi-step orchestration, with context optimization achieving a 15…

What tools did this team use?

OpenAI, Anthropic.

What results were reported?

token, inference latency, and cost reduction from ID aliasing: ~15–30%; work automatable by AI agents: thousands of hours of work in seconds (source-reported, not independently verified).

What failed first in this deployment?

How is this back office ops AI workflow structured?

User message triggers agent loop → Context manager assembles world state → Decision engine calls LLM → Tool dispatcher executes tools → LLM self-corrects on errors → Final response ends agent loop.