Skip to content
Product Agents

Search Product Agents

Start with a guide, or search a topic

    New here? Read the five guides in order

    Subscribe

    Airtable Omni & field agents

    Airtable’s agent is a column. One approval becomes a standing rule.

    Airtable’s AI writes into columns that feed formulas, automations and other agents, and reruns on its own. Read this before one checked answer stands in for approving a rule that keeps running, under an owner that changed in September.

    An Airtable Photography Brand Review interface where field agents mark campaign photos Failed with brand-check scores, and the City Wall Lean record explains a lighting issue while its status shows In progress.
    Image: Airtable recording · June 27, 2025 · 12:55

    At a glance#

    What it does

    Omni, a chat assistant, builds apps and fields and edits records on request. Field agents are AI columns that write a value for every record: classify, pull facts from documents, search the web, generate images, link related records, and read and write to 16 outside tools.

    Who it’s for

    Teams that run their work in Airtable bases. Airtable counts more than 500,000 organizations, including 80% of the Fortune 100.

    Price

    AI credits come with every plan since June 24, 2025. Free: 500 a month per editor. Team ($20 a seat a month, billed yearly): 15,000 per paid user. Business ($45): 20,000. Enterprise Scale: 25,000. Building with Omni is free. Running spends credits: a question or a web search is 10, one piece of feedback categorized is 1, a 10-page contract analyzed is 200. Extra packs from $20 a month for 10,000.

    Model

    Picked per field, and the options depend on your plan. Enterprise customers can choose OpenAI GPT models, Anthropic Claude models through Amazon Bedrock, and Meta models.

    Takes actions?

    Yes, inside the base and outside it. Omni edits records within the user’s permissions and offers an Undo per response. Field agents write cells and, since April 2026, push data to Gmail, Outlook, HubSpot, Jira and 12 other tools, with no confirmation step per record.

    Launched

    Omni and field agents on June 24, 2025, the “AI-native” relaunch. Deep Match, which links related records, February 23, 2026. Scheduled and conditional triggers by April 24, 2026. Outside-tool connectors April 2026.

    Owner / Status

    Bending Spoons agreed to buy Airtable on August 4, 2026 for $1.285 billion in enterprise value, about $2.25 billion with cash, against an $11 billion valuation in 2021. The deal closed September 4, 2026. Co-founder Howie Liu stepped down as CEO to run Hyperagent, the agent business spun out before the sale. Bending Spoons’ CEO Luca Ferrari now leads Airtable.

    Why it matters#

    Airtable rebuilt itself around AI in June 2025: Omni, a chat assistant that builds the app, and field agents, AI columns that fill a value for every record. Credits went into every plan, including Free. Fourteen months later Airtable sold to Bending Spoons for $1.285 billion in enterprise value, against an $11 billion valuation in 2021. Its founder left with the spun-out agent business.

    In Airtable the agent’s output is a column. Formulas read it, automations fire on it, other agents use it, and it can rerun when its inputs change. Approving the first value approves a rule that keeps running, and each run costs credits. This page covers the triggers, the human-edit lock, the credit meter and the warning about hidden fields.

    How it works#

    What happens when a request reaches Airtable Omni & field agents.

    1. Ask Omni

      A builder attaches a brand guidelines document and types: “Add the new brand guidelines from the attached doc.” Building costs no credits.

    2. Omni builds

      Nine seconds later five new rules sit in a Brand Guidelines table, each ticked to be included in the brand check. Omni works within the builder’s permissions, and an Undo icon appears on the response.

    3. Field agents run

      Five Brand Check fields and a Failure Reason field start on every record at once. Each cell is a model call paid in credits; analyzing a 10-page document costs about 200.

    4. Results land as data

      Every cell reads PASS or FAIL with a reason beside it. Brand Check #1 fails six assets for “Lacks half-marathon readiness”, a phrase the rule gave only as an example.

    5. The column keeps running

      Set to run on change, the field reruns when its inputs change; since April 2026 it can also run on a schedule or when a condition is met. Other fields, formulas and automations read the output. A cell a person edited is never overwritten, and stale or failed cells are marked so a rerun can target just those.

    6. Review a record

      A failed asset like City Wall Lean shows its photo, the verdict and the reason. It doesn’t show the prompt that produced the verdict or which fields the check read.

    The decisions worth studying#

    What they chose, why, and what it costs.

    1Building is free. Running is metered by the cell.

    Omni, Airtable’s chat builder, costs nothing to build with: creating apps, fields and agents spends no credits. Credits go when the work runs. Every plan has included them since June 24, 2025: 500 a month per editor on Free, 15,000 per paid user on Team, 20,000 on Business and 25,000 on Enterprise Scale, pooled across the company. Extra packs start at $20 a month for 10,000. Airtable’s own examples: a question or a web search is 10 credits, categorizing one piece of feedback is 1, analyzing a 10-page contract is 200.

    My read: this fits the column. The unit of cost is the cell, and the trigger decides how many times you pay for it. A field set to run on change spends again every time an input changes. One builder in Airtable’s community described agents “generating images and consuming credits every time I updated some fields”. Airtable’s answer, in March 2026, was more trigger options rather than a cheaper run.

    2A field reruns by itself, and since April 2026 on a schedule

    A field agent can run four ways: when a person clicks Run agent, automatically when the fields it reads change, on a schedule, or when a record meets a condition. Airtable’s help page says to treat automatic runs “much like an automation” and weigh the downstream effects, because other fields, formulas and automations read the output. Since April 2026 a field agent can also read and write to 16 outside tools, including Gmail, Outlook, HubSpot and Jira.

    Airtable’s June 2025 showcase shows how this compounds. A builder asks Omni to add brand rules from a document; five rules appear in nine seconds. One rule gives “half-marathon ready” as an example of a performance cue. Five check fields run on every record, and the first fails six assets for lacking exactly that phrase. Whoever approved the first value approved all six, and every future one.

    My read: the review step belongs before the first run, on the trigger and the inputs. Afterward there is no step left to review.

    Airtable · Field agentsProduct screenshot
    The same columns with results: most cells read PASS; Brand Check #1 reads FAIL on six rows, and each of those rows’ Failure Reason cites Concise Copy and “Lacks half-marathon readiness.”
    Six assets fail Brand Check #1 for the same reason, a phrase the rule gave only as an example. Each cell is a model call, and the column reruns when its inputs change. Airtable recording · June 27, 2025 · 11:51

    3A cell a person edited is never overwritten

    In an AI field set to generate automatically, Airtable will never automatically overwrite a cell a human edited, and that holds after the prompt or settings change. When a builder edits the prompt, Airtable asks whether to regenerate existing records, and the human-edited cells stay out of it. The grid also tracks which cells the agent modified, which are outdated and which have errors, and a bulk run can target just one of those groups.

    Airtable’s reasoning is respect for people: a correction is a decision, and a decision outranks a generated value. The cost is a column that holds three kinds of values at once. Cells from the current prompt, cells from an older prompt, and cells a person maintains. A filled cell looks the same in all three cases. The showcase record for City Wall Lean shows a verdict and a reason, but not the prompt that wrote it or the fields it read.

    My read: the lock is right, and it needs a label.

    Airtable · Record detailProduct screenshot
    Record detail for City Wall Lean: an In progress status with its tooltip, the Produced Shot w/ Copy image reading “Earthy Strength. Planet Friendly.”, and the Failure Reason “Lighting Issue - Does not prioritize natural or diffused lighting, using harsh or colored spotlights outside of intentional, stylized contexts.”
    The record pairs the photo with a verdict and a reason. Nothing on it says which prompt version wrote the value, or whether a person has since changed it. Airtable recording · June 27, 2025 · 12:55

    4The founder built a separate agent, then sold the base

    Seven months after the relaunch, on January 27, 2026, Airtable launched Superagent, a standalone research agent at $20 to $200 a month, built on DeepSky, a startup it bought in October 2025. By July 2026 the product was Hyperagent, Airtable’s agent platform; superagent.com now redirects to hyperagent.com.

    On August 4, 2026, Bending Spoons agreed to buy Airtable for $1.285 billion in enterprise value, against an $11 billion valuation in 2021. Airtable’s recurring revenue was about $480 million, growing over 20% a year. Before the agreement, the Hyperagent business was moved into a separate company. The sale closed September 4, 2026. Howie Liu wrote that he was stepping down as CEO to work full time on Hyperagent, “now an independent company”. Bending Spoons’ chief executive, Luca Ferrari, now leads Airtable.

    My read: Omni and field agents are the agents Airtable kept. They now belong to an owner known for buying Evernote, WeTransfer and Vimeo and cutting staff afterward. Judge Ferrari’s promise of heavy investment by what ships.

    What to watch#

    • New owner, known playbook

      Bending Spoons has cut staff deeply after past purchases; its CEO said moving Evernote to Europe meant “significantly reducing the size of the US-based team”. Watch Airtable’s release notes and support response times through 2027.

    • Hidden fields can surface

      Airtable warns that a hidden field referenced in a prompt can appear in the AI output that interface collaborators see. Telling the model to skip a field is weaker than leaving it out of the inputs.

    • Undo covers one Omni response

      Omni shows an Undo icon on responses where it built or changed something. How far undo reaches into field runs, downstream automations or emails a field agent sent is not documented.

    • Connected tools run as one person

      Every connector on a field agent must use the same user account, and each needs its own sign-in. Writes to Gmail or HubSpot happen per record with no confirmation, under that one account.

    What it means for your team#

    What to decide before you ship an agent like Airtable Omni & field agents.

    • Product

      Before the first value is written, show what the field reads, when it reruns, what reads its output and who sees it. Airtable now offers four triggers; make the trigger a visible choice.

    • Design

      Mark which cells a person edited and which are outdated or failed, so a reviewer can tell the agent’s values from a person’s. Airtable keeps the lock, but a filled cell looks the same either way.

    • Security

      Check each AI column’s inputs against who can see its output; a hidden source field can appear in the AI value. Connectors act as one person’s account for every record.

    • Finance

      Budget by trigger and record count. A Team seat carries 15,000 credits a month, a 10-page contract costs 200, and a column set to run on change spends again every time an input changes.

    Steal this

    1. Never overwrite a value a person corrected. Treat their edit as a lock.

    2. Make the rerun trigger a visible choice, and show which cells went stale before rerunning.

    3. Charge nothing to build and meter the run, then show the meter next to the trigger.

    Reference#

    The evidence behind this teardown: recordings, sessions, what Airtable Omni & field agents gets right and wrong, and every source.

    Walkthrough · 5 frames

    From “Add the new brand guidelines” to a failed asset in progress#

    Frames from Airtable’s LA product showcase recording, June 27, 2025. A vendor recording of the June 2025 interface, not the September 2026 product and not a live test.

    1. Step 1 of 5: The builder hands Omni the new guidelines

      Airtable’s AI Assistant panel, marked Beta: the attachment Apex Photography Guidelines.docx, the request “Add the new brand guidelines from the attached doc”, and a Thinking indicator.
      Image: Airtable recording · June 27, 2025 · 11:22

      1. The builder attaches Apex Photography Guidelines.docx and asks, in one line, to add the new brand guidelines.

      2. The June 2025 panel reads “AI Assistant”, marked Beta; the presenter calls the assistant Omni.

    2. Step 2 of 5: Five new rules join the brand check

      Airtable’s Brand Guidelines table with Brand Rule, Rule Description and Include in Brand Check columns: five earlier rules with grey ticks, then five new rules with green ticks, from Prominent Cont… to Balanced Athlete Framing. The Concise Copy rule reads “Keep copy under 25 words and include (a) the setting, (b) a single product feature, or (c) a performance cue (e.g., "half-marathon ready").”
      Image: Airtable recording · June 27, 2025 · 11:31

      1. Five new rules now sit under the original five, nine seconds later in the recording.

      2. The Concise Copy rule accepts a setting, a product feature or a performance cue; “half-marathon ready” is only an example.

      3. Each new rule arrives ticked “Include in Brand Check”.

    3. Step 3 of 5: Field agents start on every record

      An Airtable grid: every Brand Check #1 to Brand Check #5 cell reads “Analyzing files…” and every Failure Reason cell reads “Generating…”; the toolbar shows 21 hidden fields, Filter, Group, Sort, Color and Share and sync.
      Image: Airtable recording · June 27, 2025 · 11:41

      1. Five Brand Check fields and a Failure Reason field start on every record at once.

      2. The view hides 21 fields, and the grid doesn’t show which fields each check reads.

    4. Step 4 of 5: The first check fails six visible assets

      The same columns with results: most cells read PASS; Brand Check #1 reads FAIL on six rows, and each of those rows’ Failure Reason cites Concise Copy and “Lacks half-marathon readiness.”
      Image: Airtable recording · June 27, 2025 · 11:51

      1. Brand Check #1 fails six of the visible assets, while checks #2 to #5 pass them.

      2. Each Failure Reason cites Concise Copy and “Lacks half-marathon readiness”, though the rule offered that wording as one example.

    5. Step 5 of 5: A failed asset sits in progress

      Record detail for City Wall Lean: an In progress status with its tooltip, the Produced Shot w/ Copy image reading “Earthy Strength. Planet Friendly.”, and the Failure Reason “Lighting Issue - Does not prioritize natural or diffused lighting, using harsh or colored spotlights outside of intentional, stylized contexts.”
      Image: Airtable recording · June 27, 2025 · 12:55

      1. City Wall Lean fails on lighting instead: harsh or colored spotlights outside intentional, stylized contexts.

      2. The record reads “In progress”; the later frames show no owner, fix or recheck for City Wall Lean.

      3. The record pairs the produced shot with a verdict, but not with the prompt or the fields the check read.

    Sources#

    1. Airtable AI billingsupport.airtable.com
    2. Using Airtable AI in fieldssupport.airtable.com
    3. Using Omni AI in Airtablesupport.airtable.com
    4. Airtable pricingairtable.com

    Sources checked September 28, 2026.

    Next teardownMicrosoft Excel agents

    Microsoft has two Excel agents: Copilot inside Excel, which edits the open workbook, and Excel Agent in Copilot Chat, which creates a new workbook file.