If you run Hightouch, an agent will soon wire up the tools its catalog doesn't cover instead of your data engineers, and nobody has said who signs off on what it writes into production. Hightouch announced on October 9, 2026 that it has acquired Strata, a two-year-old startup whose agent builds and maintains custom API integrations from a plain-language description. Terms were not disclosed. The announcement doesn't cover Strata's standalone customers, pricing, or how an agent-built integration is tested and approved before it starts moving customer data.
Today, when a destination outside Hightouch's native catalog changes its API, your own engineers fix the integration. Hightouch wants its agent to do that work instead, which changes who is accountable when a sync breaks.
What Hightouch Actually Bought
Hightouch bought an agent that turns a sentence into a working integration and then keeps it running. In Hightouch's own description, Strata is "an agent-driven platform for building and maintaining custom API integrations without code": a team describes the data it wants moved and where, and the agent maps fields and events, configures transformations, generates the integration, and keeps it running as systems change.
The founders know the territory. Kelsey Mullaney, Julian Mclain and Ilya Brin founded Strata in 2024 after meeting as early employees at Iterable, the marketing automation platform, according to the Business Wire release. Mullaney, Strata's CEO, joins Hightouch to help lead revenue programs and go-to-market strategy. That is a sales and go-to-market job, which suggests Hightouch wants Strata's founder selling integrations as well as building them.
Strata's own site, connectstrata.com, pitches bi-directional flows (webhooks in, jobs run against APIs out), historical backfills with "a suite of dry-run and testing tools," and a promise of "integrations that don't break as your systems evolve." It shows three named testimonials and no customer logos, pricing, or security certifications. As of October 11, the site carries no notice about the acquisition.
Hightouch's stated reason is speed. Co-founder and CTO Joshua Curl said in the release that "marketers should be able to use their customer data without waiting on engineering for every new campaign or connection." The company also frames the deal as feeding its own agents: Strata helps it connect more of the tools marketing teams use so its AI agents have the context they need.
Why the Connector Catalog Was Never Going to Be Enough
A connector catalog always ends somewhere, and Hightouch's has been growing for years without closing the gap. In 2022, Hightouch claimed "100+ native destinations" when it launched its custom destination toolkit. Its HTTP Request documentation now cites "250+ native destinations", while its pricing page FAQ says "300+" integrations. Whichever count you use, the long tail of martech (the regional SMS provider, the loyalty platform your acquired brand runs, the internal API your commerce team stood up last year) sits outside it.
For that tail, Hightouch has offered a do-it-yourself path since 2022: HTTP requests, serverless functions, message queues and databases. The HTTP Request destination lets you template a payload, set auth, and tune rate limits. The page is explicit that it is "strictly prohibited" to use it to replicate a native destination, so the custom path exists only for tools Hightouch doesn't cover.
Someone on your side builds and owns those custom destinations. The 2022 toolkit announcement doesn't say who maintains them when the third-party API changes, and in practice the answer has been whoever built them. Strata is Hightouch's attempt to move that maintenance burden to an agent it runs.
The strongest case for the deal is that the maintenance burden is real and expensive. Fivetran, which sells connectors for a living, wrote in a June 30, 2026 post that AI-generated connectors leave "schema drift, validation, API version changes, retry logic, authentication failures" on your backlog, and that AI "doesn't change who gets paged when the pipeline fails." If an agent at Hightouch absorbs that pager duty for the long tail, your data engineers get time back.
What Changes When an Agent Maintains the Integration
The shift is from a human-authored integration you can read to an agent-maintained one that changes when the agent decides a destination has changed. Building is the easy half; "keeps it running as systems change" is the part that needs a control model, and none has been published.
Look at how the rest of the market handles the same problem. Fivetran's coding-agent plugins scaffold a connector, run a local test and try to auto-fix errors, and its documentation still says: "We recommend reviewing AI-generated connector code before deploying to production." Airbyte's AI Assistant for its Connector Builder is labelled beta, says "it won't always be perfect," and tells users "you should still review the output." Both vendors put a human between the generated connector and production, and both tools build source connectors that pull data into a warehouse.
Hightouch's direction is the riskier one. Reverse ETL writes to operational systems: the email platform that sends to customers, the ad account that spends money, the CRM your sales team works from. A wrong field mapping in an inbound pipeline produces a bad dashboard; in a reverse ETL sync it could send an offer to a suppression list or overwrite a consent flag.
Hightouch's existing error handling won't catch that class of mistake. Its HTTP Request docs treat any 400- or 500-level response as an error and retry it until it succeeds. The payload editor has a Preview tab that renders a template against a sample row; the page describes no live test request. A destination that accepts a mis-mapped payload returns success, so a silent mapping change made by an agent would pass every check those docs describe.
What Hasn't Been Said, and Who Should Care
Strata's own customers are in the weakest position. The announcement says nothing about whether the standalone Strata product continues, for how long, or at what price, and Strata's site lists no pricing to hold Hightouch to. Recent small-target deals we covered went badly for standalone users: Juicebox gave Fetcher's customers until October 16, and Mistral's Pimento deal left 180 brands without a roadmap. Strata's integrations need credentials to your marketing systems, and the problem from Adobe's Rilo acquisition applies here too: a grant can outlive the product that requested it.
Hightouch customers have a pricing question. Hightouch's pricing page publishes no dollar figures; both paid plans route to "Get a demo," and the Composable CDP plan lists "Data Agents" among its inclusions. Nothing says whether agent-built integrations will be bundled, metered, or sold as an add-on. Hightouch raised $150M at a $2.75B valuation in April, with more than $100M in estimated annual recurring revenue, and the same report says it does not plan to be profitable this year or next. A company in that position has reason to turn a capability like this into a line item. We covered the strategy behind that raise in Hightouch's $2.75B warehouse play.
Teams that built their own custom destinations should decide now whether they want an agent touching them at all. If your HTTP Request destinations are hand-built and stable, you lose nothing by waiting for the review model before opting in.
What to Do Before You Let an Agent Write to Production
Treat an agent-maintained integration as a change to a production system that you didn't author, and apply the controls you'd apply to any such change.
This Week:
- Inventory your custom destinations. List every HTTP Request, serverless function and webhook destination in your Hightouch workspace, who built it, and what it writes. That list is your exposure if agent maintenance becomes the default.
- If you use Strata directly, ask for continuity in writing: how long the standalone product runs, what it will cost, and how you export integration definitions. Then list every system Strata holds a credential for, so you can revoke grants on the day it shuts down.
- Set up row and sync error alerts on every custom destination, which Hightouch's own docs already recommend. They won't catch a silent mis-map, but they catch the loud failures.
Before You Opt In:
- Ask Hightouch four questions and get the answers on paper. Does the agent change an integration's field mapping without human approval? Is every change versioned with a diff you can read? Can you roll back to the prior version in one step? Is there a dry run against the live destination before a change goes active?
- Require an approval gate for anything that writes to consent, suppression or spend. Our guide to human-in-the-loop approval gates covers how to keep that gate from becoming a rubber stamp.
- Run a reconciliation check. For one agent-built integration, compare a sample of rows in the destination against the source model every week, because a 200 response only tells you the API accepted the payload.
Before Renewal:
- Get agent-built integrations priced in the order form, with a cap, before they show up as usage.
- Add change-of-control and exit terms that cover export of integration definitions. Our breakdown of AI vendor exit clauses has the language to start from.
If your marketing team already routes long-tail tools through Zapier or Workato, keep those flows until Hightouch publishes how its agent's changes are reviewed.
The Bigger Picture
Connector vendors have long worked on one split: the vendor builds the popular connectors, and the customer builds and babysits the rest. Hightouch's own 2022 toolkit was a better set of tools for the customer's half. Fivetran, Airbyte and now Hightouch are using agents to shrink that half, and the first two have documented a human review step. Hightouch hasn't yet, and its agent will be writing to the systems that send email to your customers and spend your ad budget.
Ask Hightouch for its review and rollback model before your first agent-built destination goes live.
Continue Reading
- Hightouch's $2.75B Warehouse Play Just Broke Salesforce CDP
- Adobe Took Rilo's Team. Its Slack Token Never Expires.
- Fetcher Shuts Down October 16 After Juicebox Buys Its Assets
- Mistral Bought Pimento for Vibe, Leaving 180 Brands No Roadmap
- AI Vendor Exit Clauses: Standard Terms Give 0 to 90 Days to Get Out
- Human-in-the-Loop for AI Agents: Most Approval Gates Rubber-Stamp