Factory, the company behind the Droid autonomous coding agents, said on October 30, 2026 that it is acquiring Typedef, a startup that builds what it calls a data context layer. The stated aim, per Factory's news post and its account on X: give Droids, and the engineers working with them, "clearer lineage and evidence to plan migrations, debug pipelines, and refactor models" (Factory news; Typedef). Typedef's site now carries a banner reading "Factory acquires Typedef."
The deal terms were not in the material we could read, so we do not report a price. What is checkable is what Typedef sells, how Factory describes the plan, and one headline number: Typedef's claim that agents succeed 82% of the time with its layer versus 46% without.
TL;DR
| Question | Answer |
|---|---|
| What happened? | Factory is acquiring Typedef, announced October 30, 2026 |
| What does Typedef make? | A live data context layer compiled from your code, covering definitions, pipelines, dashboards and systems |
| What will Factory do with it? | Give Droids and engineers lineage and evidence for migrations, pipeline debugging and model refactoring |
| Price or terms | Not disclosed in what we found |
| Headline claim | Typedef says 82% task success with the layer versus 46% without |
| Independent check | None; the figure is on Typedef's website with no linked dataset |
| Security posture claimed | SOC 2 Type II, GDPR, VPC deployment, no data egress |
What a data context layer is
A green thread following a cream seed through three stages, a picture of data lineage context for AI coding agents
Coding agents are good at reading code but weak at knowing what code touches. In data engineering this shows up constantly: change a staging model and a dashboard three hops downstream quietly breaks. Typedef's pitch is a layer that records those dependencies so the agent does not have to guess.
Per Typedef's site, the layer is "a semantic ontology of every definition, pipeline, dashboard, and system across your data platform, compiled from your code, always current, always traceable." It lists use cases:
- Blast radius before you merge: change a staging model and see every downstream report, metric and dashboard affected before the pull request lands.
- Root cause in minutes: trace a broken pipeline to the failing node and the upstream change behind it.
- Metric reconciliation: trace each field from source report to BI layer and see where two systems' numbers diverge.
- Safe change plans: an implementation plan for each layer with verification steps and risk analysis.
- Certified metrics: compare how a metric is computed across models, including filters, grain and date logic, and flag ones that are not safe to sum.
- Trustworthy agentic analytics: back BI agents with a shared model of the stack so answers are explainable.
Why Factory wants it
Factory's own news feed shows the direction. In the weeks before the acquisition it announced a Databricks collaboration on software and data modernization (October 19), a self-hosted sandbox orchestrator called OpenCompute (October 14), a Droid 1.0 harness (October 29) and an earlier acquisition, Mentlio, for token allocation (October 9). The Typedef deal fits a pattern: enterprise customers want agents that can migrate and refactor data systems, and those tasks fail when the agent lacks the map of what depends on what.
The announced uses are specific: plan migrations, debug pipelines and refactor models.
For a sense of how much the harness around a model matters, see our look at what a harness buys you in tokens and why the harness is the company.
The 82% vs 46% claim: what it is and is not
Typedef's homepage shows two numbers under "Task Success Rate": 82% with its data context layer and 46% without. It defines the metric as "the percentage of real-world data engineering tasks completed successfully against production-grade acceptance criteria," tested across impact analysis, root-cause debugging and model refactoring, and says "agents with full lineage context dramatically outperform those without it."
What we can confirm: the numbers are on the page, and the task categories are named. What we cannot confirm:
- Task set. No dataset, task count or repository list is linked from the page.
- Agent and model. The page does not say which agent or model produced either score.
- Baseline fairness. "Without" could mean an agent given only code, or something weaker. Not stated.
- Grading. "Production-grade acceptance criteria" is not described, nor who graded.
- Independence. It is the vendor's own test, published on a sales page.
That does not make it wrong. A lineage map plausibly helps on these tasks, and a gap of this size is not implausible for impact analysis, where a code-only agent simply cannot see dashboards. But a vendor figure on a marketing page should be reproduced before you plan around it. Compare how research benchmarks publish their setups, for example the Co-Design Bench release, which links its code, dataset and grading protocol, or FrontierCode's maintainer-merge benchmark.
What this means for teams using coding agents
If you run dbt or a similar stack, the claim is easy to test on your own work. Pick ten recent changes where you know the downstream impact, ask an agent for blast radius with and without a lineage source, and count the misses. Do the same for three past pipeline failures.
If you evaluate agents, note the shift from model comparisons to context comparisons. Cost control is part of that story too, as in LangChain's model router results and Databricks on managing AI coding costs at scale.
If you worry about data exposure, Typedef claims a VPC deployment, SSO, role-based access, audit trails and "no data egress," plus SOC 2 Type II and GDPR alignment. How that carries over once it sits inside Factory is not described in what we read; ask for the post-acquisition architecture and data-handling terms. Our piece on treating agents as insiders covers the control questions to ask.
What people are asking
Is this an acqui-hire or a product acquisition?
Factory's post says Typedef "is joining Factory to bring deeper data context to autonomous software engineering," and its X post welcomes Typedef's founders by name. That wording covers both the team and the technology, but it does not say whether Typedef's standalone product continues. Typedef's site still presents the product and a demo booking button, with the acquisition banner on top. Until Factory says otherwise, assume existing Typedef customers should ask about continuity and support.
How does this relate to Factory's other recent moves?
The company's October news shows a push toward enterprise data and deployment: a Databricks collaboration, self-hosted sandboxes, a new Droid harness and the earlier Mentlio acquisition aimed at token allocation. Typedef adds the missing piece for data work, which is knowing what a change affects. None of those posts, as far as we read, gives usage numbers, so we cannot say how widely any of it is deployed.
Does lineage replace tests and review?
No. Lineage helps an agent see blast radius and trace failures, but it does not decide whether a metric definition is correct. Typedef's own list of use cases, such as certified metrics and reconciliation, points to review workflows rather than autonomous sign-off. Treat it as an input that makes human review faster and more targeted.
How to test whether a context layer helps your team
A fair test is small and cheap to run, and does not need Factory or Typedef:
- Pick the tasks. Choose around ten past changes to models or pipelines where you already know the downstream effects, plus three incidents where you know the root cause.
- Define success first. Write acceptance criteria before running anything, for example "lists every downstream dashboard" or "names the upstream commit," so that grading does not drift.
- Run two arms. Give the same agent and model the repository alone in one arm and the repository plus a lineage export in the other.
- Count misses, not impressions. Record missed dependencies and wrong causes, along with time and token cost for each arm.
- Repeat on a second repo. Context layers can overfit to a clean project; a messy one shows more.
If the gap on your own tasks looks like 82% versus 46%, the vendor number holds for you. If it is closer to ten points, the layer may still be worth it but the sales page overstated the effect for your stack. Either way you will have a number you can defend, which is more than the page offers.
What we do not know yet
- The price, structure and team that is joining.
- Whether Typedef will remain available as a standalone product.
- When Droids will gain features built on the layer, and on which plans.
- Which warehouses and transformation tools are supported beyond dbt.
- Whether any independent evaluation of the 82% and 46% figures will appear.
We will update this post if Factory or Typedef publish terms or an evaluation.
Related reading
- What does a harness buy? Tokens, mostly
- The harness is the company
- Co-Design Bench: real app edits for coding agents
- FrontierCode 1.1 and maintainer merges
- LangChain Open SWE model router
- Databricks: managing AI coding costs
- Treating agents as insiders
Details reflect Factory's news page and Typedef's website as of October 30, 2026; terms and features may change.