DAGWorks Inc.
Winter 2023AcquiredOpen source tools & services for reliable AI Agents & AI Applications
Pressure-test this opportunity
Explore the risks and possibilities with a prompt for ChatGPT, Claude, or your agent.
Open source tools & services for reliable AI Agents & AI Applications
Explore the risks and possibilities with a prompt for ChatGPT, Claude, or your agent.
DAGWorks was a two-person YC W23 company built around a sharp engineering idea: ordinary Python functions could declare a dataflow graph, making machine-learning and data pipelines easier to test, inspect, and move between execution systems. Hamilton supplied that open-source foundation; Burr later applied explicit state machines, persistence, and telemetry to AI agents. The commercial layer sold hosted observability and collaboration around both projects.[1][2]
This is an acquisition story, not a conventional failure. Salesforce acquired DAGWorks in early 2025, and co-founder Elijah ben Izzy moved into Agentforce work on Agent Script and determinism.[3] The more instructive tension is that DAGWorks created credible open-source primitives, but had to commercialize a broad platform while its underlying categories were still shifting. Salesforce could capture the team’s expertise; Apache could steward the code. Public evidence does not establish that the hosted business achieved independent scale.
Stefan Krawczyk and Elijah ben Izzy founded DAGWorks in 2022 and joined YC’s Winter 2023 batch. The available primary sources identify them and the product, but do not establish how they met, their precise division of responsibilities, or a detailed financing history. No reliable founder interview surfaced with the two verbatim origin quotes requested by the canonical format, so this report does not manufacture them.
Their initial thesis came from a familiar problem in production machine learning: teams wrote pipelines as procedural scripts, then accumulated hidden dependencies, brittle orchestration code, and debugging work. Hamilton inverted the model. Developers wrote small Python functions whose names, inputs, and outputs described a graph. That graph could be visualized, tested, and executed without binding the business logic to Airflow, Dagster, or another orchestrator.[4]
The company paired the open-source library with a proprietary control plane. The platform tracked executions, catalogued artifacts, exposed lineage, and helped teams debug failures. In its YC launch, DAGWorks positioned Hamilton as the expression layer rather than another scheduler, and invited teams into a private beta for the commercial product.[2]
Burr expanded the ambition. Instead of data assets, Burr modeled an application as actions and transitions over explicit state. Persistence, replay, human intervention, and telemetry made it relevant to chatbots, RAG systems, simulations, and agents.[5] The move was technically coherent: both products turned implicit Python behavior into inspectable graphs. Commercially, however, it widened the market story from data and ML workflows into a fast-changing agent framework category.
Hamilton treated a Python function as a node and its parameters as upstream dependencies. A developer could write transformation logic as plain functions, ask Hamilton to assemble the graph, and run only the outputs needed for a job. This made lineage visible and unit testing natural. It also separated computation from orchestration: teams could keep their Python model while changing where or how it ran.[4]
The commercial platform added execution history, a data catalog, observability, debugging, and integrations. The pitch resembled “dbt for Python,” but covered ML features and application logic that lived outside SQL. A browser playground lowered the trial cost, while the open-source package made adoption possible without a platform purchase.[2]
Burr used the same inspectability principle for stateful applications. Developers declared actions, transitions, and state rather than burying control flow inside loops and callbacks. The runtime could persist state, resume work, display telemetry, and support human approval. That architecture anticipated a real agent-engineering problem: generated behavior is difficult to operate unless teams can reconstruct what happened and why.[10]
Hamilton and Burr therefore shared a product philosophy, but not necessarily a single buyer or budget. Hamilton addressed data and ML platform teams. Burr entered a crowded agent-runtime market alongside LangGraph, Temporal, and LangChain. The common commercial product was observability, yet each ecosystem brought different integrations, workflows, and sales language.
Hamilton targeted Python-heavy data, machine-learning, and platform teams that wanted testable transformations without replacing their orchestrator. Its repository documents production references across Stitch Fix, UK Government Digital Services, IBM, Opendoor, LexisNexis, Adobe, the Federal Reserve Board, Joby, and EquipmentShare. These references show adoption, not paid-customer or revenue totals.[4]
Burr targeted teams shipping stateful AI applications, especially agents and RAG systems that required persistence, replay, telemetry, and human intervention.[5]
DAGWorks did not publish a defensible market-size calculation, revenue figure, customer count, or annual recurring revenue. Public pricing still advertises a 14-day hosted Hamilton Team trial and labels Burr Cloud pricing “Coming Soon,” but a live page does not prove that either service remains operational after the acquisition.[11]
Hamilton’s strongest positioning choice was to avoid competing head-on with orchestration. Airflow, Dagster, and Prefect decide when and where jobs run; Hamilton describes the transformations and their dependencies. That distinction let teams adopt it inside existing infrastructure. Its weakness was budget ownership: an expression layer can create engineering value while remaining difficult to monetize separately from the orchestrator, warehouse, or ML platform.
Burr faced a more fluid contest. LangGraph and LangChain owned attention around agent construction, while Temporal brought a mature durable-execution model. Burr differentiated on explicit state and inspectability, but framework adoption tends to pull observability along with it. A small vendor therefore had to win both developer mindshare and a commercial control-plane sale.
The acquisition suggests a different form of validation. Salesforce placed ben Izzy on Agentforce Agent Script and determinism, work closely aligned with Burr’s emphasis on explicit, reproducible execution.[3] That connection supports an inference that DAGWorks’ technical expertise mattered to Salesforce. No public Salesforce announcement explains the deal rationale, so it should not be presented as confirmed transaction logic.
DAGWorks used open source for adoption and sold hosted team capabilities around observability, lineage, versioning, and collaboration. The company’s 2023 retrospective says it reached its first paying customers, but gives no revenue, contract size, retention, or margin data.[6] Public pricing later separated free use from Team and Enterprise plans, with Burr Cloud not yet generally priced.[11]
No primary financing announcement or transaction consideration was located. Secondary estimates of roughly $500,000 are too weak to use as fact. With only two founders listed by YC, the company was likely capital-efficient, but any burn-rate estimate would be guesswork without payroll, infrastructure spend, contractor costs, or financing dates.
DAGWorks’ 2023 review reported 1,100 stars, 58 Hamilton releases, at least three Hacker News front-page appearances, a TLDR newsletter placement reaching more than one million readers, 125% growth in Hamilton’s Slack membership, conference talks, Dataflow Hub, and initial paying customers.[6] The current Hamilton repository also lists a meaningful set of production users.[4]
Those signals establish developer interest and real usage. They do not establish commercial scale. Missing figures include active installations, cloud conversion, ARR, retention, enterprise contract values, and the share of users attributable to Hamilton versus Burr.
Hamilton earned adoption by fitting into existing Python stacks. The same portability that helped distribution constrained capture: users could keep the library and use their existing scheduler, logs, and internal tooling. DAGWorks responded with catalog, lineage, and debugging features, then reported its first paying customers in 2023.[6] Public evidence does not show that conversion reached a durable scale before the Salesforce acquisition.
The non-obvious mechanism is budget fragmentation. Hamilton improved code quality at the expression layer, while buyers often assigned spending to orchestration, warehouses, model platforms, or observability suites. A useful developer primitive could spread widely and still leave its creator negotiating for a narrow slice of an already allocated platform budget.
Burr gave DAGWorks a credible answer to the rise of agents. Explicit state, persistence, replay, and telemetry addressed problems that become acute once LLM applications leave demos.[5] The move also placed a two-person company against agent frameworks with stronger distribution and against durable-execution vendors with mature enterprise positions.
The attempted remedy was technically logical: unite dataflow and application state under inspectable graphs, then sell one operational layer. The cost was a wider integration and education burden. Hamilton’s buyer, Burr’s buyer, and a Salesforce Agentforce team may all value determinism, but they encounter the problem through different budgets and workflows.
Salesforce acquired DAGWorks in early 2025. Krawczyk’s May 28 announcement and DAGWorks’ About page confirm the combination, while Salesforce’s biography confirms ben Izzy joined through the acquisition and now works on Agentforce.[8][12][3] The price, structure, signing date, close date, investor return, customer treatment, and Krawczyk’s exact role were not disclosed.
Calling the deal an acqui-hire would exceed the evidence. A narrower inference is defensible: Salesforce found value in a team working on deterministic agent execution, while the standalone hosted platform’s economics remain unproven in public.
Hamilton entered Apache incubation on April 12, 2025, and Burr followed on May 24. Their repositories now identify both as Apache incubating projects, and 2026 notices show Burr assets moving away from DAGWorks-owned URLs.[4][5][9]
Donation is a sound continuity outcome for users and contributors. It also means a future commercial vendor cannot rely on exclusive ownership of the core projects. The hosted Hamilton page remains online and Burr Cloud remains labeled as forthcoming, but their current operating status is uncertain.[11]
The strongest counter-narrative is that DAGWorks did not need to become a large standalone platform to succeed. It built respected infrastructure, found real production use, and landed inside a company with a massive agent distribution surface. That is a legitimate outcome. The unresolved question is whether Salesforce bought a growing business, strategic code and expertise, or some combination; the public record cannot distinguish among them.