How Legora runs 2,000 parallel agent evals on Modal sandboxes. Legora is an agentic system that supports lawyers in research, review, and drafting across complex matters. In early 2026, as rapid gains in model capabilities made longer and more autonomous legal work possible, Legora began replacing its existing AI assistant with a full-fledged agent. To test a legal agent that works for hours across thousands of documents, Legora boots its entire application, pinned to a specific commit, inside a separate Modal Sandbox for every eval run. To read more about Legora's full evals system, check out the blog they wrote! Rebuilding evals for an autonomous agent. The shift from assistant to agent exposed two major shortcomings in Legora's existing evals framework: scale and environment fidelity. The infrastructure challenge was to create thousands of isolated Legora environments in parallel, restore the right world into each one, and let the agent complete ambitious, long-running tasks. The first problem was scalability. The assistant had been designed for smaller tasks in environments containing a few documents. The new agent needed to run for hours inside a realistic project containing thousands of documents and gigabytes of data. Where the old eval system could run on a laptop, the new one needed to run in the cloud. The even harder problem was environment fidelity. Legora's agent spans their whole product: a headless DOCX editor, self-hosted Turbopuffer instances, PDF conversion, Postgres and authorization databases, blob storage, and dynamically created code-execution sandboxes. Modal Sandboxes provide complex environments at scale. Today, each eval run executes in its own Modal Sandbox, loaded with a custom image and filesystem. While Legora evaluated multiple cloud-managed sandbox providers to replace local execution, they chose Modal for scalability, low startup latency, and support for complex environments. Each experiment kicks off 2,000 runs or more, each of which runs concurrently in a Modal Sandbox. Engineers on Legora's team are able to kick off runs from their laptops to measure the impact of changes to their agent. To manage storage for each run, Legora uses Modal Volumes. Volumes make large document collections available across Sandboxes and let the team snapshot world state, deduplicating shared setup while ensuring that repeated runs begin from the same documents, database state, permissions, and blobs. The team also cached Docker images in Volumes to speed up the Docker Compose step, with plans to move that cache to filesystem snapshots. Building the large image for Legora's full application initially led to slow eval runs. Legora addressed the bottleneck with an hourly pipeline that builds the image and pre-warms Sandboxes, reducing the reset-to-commit step from a full application build to only the delta since the previous build. A separate hourly job syncs values from Legora's key vault into Modal Secrets, while frequent secret rotation, least-privilege access, dependency-age requirements, and Modal's domain allowlisting help secure the environment. Moving from gVisor to Modal VM Sandboxes. As an early user of Modal VM Sandboxes, Legora was instead able to run Sandboxes backed by a full virtual machine. With a real Linux kernel in every Sandbox, Docker behaved as it would on a normal host. While before they had to implement workarounds to get full Docker functionality, VM Sandboxes enabled them to simplify their setup. Creating a self-improving agent loop. The new evals platform is now Legora's primary instrument for measuring agent quality. It enabled the company to run BAR, its internal benchmark of realistic, long-horizon legal tasks, and compare foundation models across quality, cost, and latency. Those results now inform production model selection as well as decisions about the design of the agent harness. The platform also changed how the engineering team works. After wrapping the evals application in a CLI and a skill, Legora enabled agents to create cases, launch experiments, and analyze results autonomously; most engineers now interact with evals through agents. Given a metric or problem, experiment tools, and a budget, an agent can hill-climb the harness until the metric improves - an approach Legora has applied to tool search, PowerPoint generation, and reproducing research papers to make the harness more cost-efficient. For example, Legora was able to cut the cost of their agent by 30% through smarter compaction and a custom tool search implementation. While the engineering team managed the process, agents were able to work via auto-research to validate and improve their own results.

Modal becomes latest decacorn after $750M raise to build cloud infrastructure for AI agents. October 15, 2026 * Modal raised $750M at a $15.75B valuation, led by Accel, with Kleiner Perkins joining. * The valuation is more than three times the $4.65B set in its May 21 Series C. * Its run rate passed $1B, and its Sandboxes product alone reached $300M. Erik Bernhardsson helped build Spotify's music recommendations, then spent six years as CTO of Better.com, and came away convinced the cloud held developers back. On October 14, the company he founded in 2021, Modal, said it had raised $750M at a $15.75B valuation, led by Accel, less than five months after it was valued at $4.65B. Kleiner Perkins joined the round, and all of Modal's existing investors took part, including Menlo Ventures, General Catalyst, Redpoint, Lux Capital, and Bain Capital Ventures. Modal said usage has tripled since its Series C, has grown more than 15x year on year, and its annualised run rate has crossed $1B. Bernhardsson's pitch starts with scarce graphics processing units (GPUs) and large AI models that are slow to load. Developers can work around both on older tools, he wrote, "but you end up in YAML hell and GPU procurement purgatory." The sandbox business behind the run rate. Modal, founded in 2021, pools computing capacity from more than 20 cloud providers into one network and lets developers run code on it without managing servers. It built its own file system, container runtime, scheduler, and networking to do that, and it does not run its own data centres. The sharpest growth is in Sandboxes. Modal launched the product in 2023, when developers began using large language models to write code that needed somewhere isolated to run. For example, a coding assistant can test a script it just wrote inside a sandbox, away from the customer's own systems. Sandboxes have since reached a $300M run rate, just under a third of the total, with close to 3 billion sandboxes launched and capacity for 1 million running at once. Customers include DoorDash, whose AI platform team uses Modal to move apps from experiment to production, Cognition, Ramp, Linear, and Hebbia, which build background agents on it. Runway generates video there, while Physical Intelligence, Generalist AI, and 1X train robots. Modal says TypeSafe AI's Jev, which it calls one of the fastest-growing products ever, served 1 trillion tokens, the chunks of text a model reads and writes, on the platform in its first three days. Bernhardsson, who created the open-source Annoy library for search and recommendations, is chief executive. Co-founder Akshat Bubna, an early engineer at Scale AI, is chief technology officer. Modal has more than 200 staff across New York, San Francisco, and Stockholm, up from more than 120 at the time of the Series C. How fast the price moved. Modal was valued at $1.1B in September 2025, when it raised an $87M Series B led by Lux Capital. Its $355M Series C followed on May 21 at $4.65B, led by General Catalyst and Redpoint. The Series D at $15.75B is more than 14 times the Series B figure in just over a year. The Series C came in two tranches, Reuters reported: the first at a $2.5B valuation, the second at $4.65B once more investors came calling. Accel and Menlo Ventures were in the second tranche, so Accel went from first-time investor at $4.65B to leading this round at $15.75B. Modal has raised roughly $1.2B across its disclosed rounds, including a seed round led by Amplify Partners and a $16M Series A in October 2023. Two investors weighed in on X. Mamoon Hamid, a partner at Kleiner Perkins, wrote that every serious AI team his firm talks to eventually says the same thing: "We run it on Modal." Lux Capital partner Grace Isford, whose firm led the Series B, wrote that Modal "radically reimagined a seamless developer experience" from the start. A crowded race for inference. The money is chasing a market that has changed shape. Gartner forecasts worldwide spending on AI-optimised infrastructure-as-a-service of $42B in 2026, up 96%, and expects inference, the running of trained models, to account for $23.3B of it, ahead of $19B for training. Rivals are raising just as fast. Baseten took $300M at a $5B valuation in January, then a $1.5B Series F at $13B on June 22. Bloomberg reported on September 23 that Baseten and Modal were both in talks to at least double their valuations, with Baseten's talks at $26B. Together AI raised $800M at $8.3B on July 1, led by Aramco Ventures, and said its annual bookings passed $1.15B in the prior quarter. Fireworks AI raised $1.505B at $17.5B on July 16 and reports a run rate above $1B. Modal's valuation is no more than about 16 times its run rate. Fireworks sits at no more than about 17.5 times, and Together AI at roughly seven times its bookings, though the two measures are not the same. While Together AI and Fireworks centre on running and customising open models, Modal describes itself as a set of building blocks, from functions to Sandboxes to training, that developers combine themselves. The number to watch is Sandboxes. If AI agents keep writing code that needs a safe place to run, a product launched in 2023 could carry more of Modal's valuation than GPU rental ever did. If the biggest clouds or the model makers build their own sandboxes, a multiple near 16 will be hard to defend.

Modal raises $750M Series D for AI execution infrastructure. The engineers at Modal took a central database out of the route an agent follows to get a computer. For the New York AI infrastructure company, co-founded by CEO Erik Bernhardsson and CTO Akshat Bubna, that scheduling decision helps explain the business behind its latest financing: developers can ask for an environment through code, while Modal takes responsibility for making the request work across a large fleet. Modal announced a $750M Series D on October 14, 2026, led by Accel at a $15.75B post-money valuation. Kleiner Perkins joined the round alongside existing investors. The financing supports a platform spanning model training, inference and isolated environments where agents can execute code, giving Modal several routes into the infrastructure budget of an AI application. The company reports crossing $1B in annualized run rate, including a $300M pace for its Sandboxes product. Those figures describe the current pace of the business, rather than revenue already earned over a full year. The interesting commercial development is how much of an application's working environment now travels with the agent. What Modal's Series D adds to its funding history. Accel's investment account confirms that the firm led the $750M financing, deepening its earlier investment. Modal names Menlo Ventures, General Catalyst, Redpoint, Lux Capital, Amplify Partners and Bain Capital Ventures among the existing investors participating. The announcement establishes the round and its post-money valuation without disclosing individual check sizes or a detailed spending budget. Modal's previous Series C announcement, dated May 21, 2026, disclosed $355M at a $4.65B post-money valuation. General Catalyst and Redpoint led that round, while Menlo, Bain Capital Ventures and Accel joined as new investors. Accel has moved from joining the syndicate to leading the next financing in roughly 5 months, a concrete change in its financial relationship with the company. Modal now says it has raised over $1.2B in total. The company reports that usage has tripled since the Series C and that growth has exceeded 15x year over year; those operating figures remain company-reported. The capital provides resources for continued infrastructure and product development as customers ask the same supplier to support a wider range of work. The announcement gives no allocation percentages, infrastructure purchase commitments or separate acquisition budget, so the financing should be read alongside the products Modal is already putting into customers' hands. Why an agent needs its development environment. An agent assigned a software task may have to run the application, inspect a database and test a change against several connected services. Modal's Sandbox documentation describes environments for executing arbitrary user or agent code, with configurable images, storage and commands. That places the runtime inside the quality of the result: the environment determines what the agent can actually execute and observe. Modal made VM Sandboxes generally available on October 1. These environments supply a Linux kernel and support workloads needing Docker, local databases and capabilities that a lighter container environment may not provide. Modal keeps both gVisor and VM runtimes available; the right choice depends on the workload. Its documentation currently limits GPU Sandboxes to gVisor, a useful distinction for teams combining code execution with accelerated compute. The customer examples give the product a recognizable place in an engineering organization. Modal says Linear's Coding Sessions put a user's development environment inside a VM Sandbox. Legora uses the environments to evaluate legal agents against an application that includes Postgres and a document editor. These are vendor-published deployment accounts, but they make the purchasing decision tangible. A team choosing an agent environment is also choosing how much of its existing application can accompany the software into the task, and how much setup its own engineers must continue to maintain. The engineering hidden behind a simple compute request. Modal's sandbox scaling account describes why the company rebuilt its scheduling system. Central coordination and writes to a shared database had become difficult to carry through very high rates of environment creation. The redesign lets scheduling servers use cached state and contact workers directly, with much of the durable recordkeeping handled asynchronously. In Modal's account, each worker becomes the authority for its own state, and scheduling behaves more like distributing requests across available machines. This is a specific engineering trade-off behind the apparent simplicity of asking for another Sandbox. It also helps explain why the founders invested in infrastructure below the API: a developer's short request can trigger a substantial amount of coordination that the customer never sees. Modal reports demonstrating 1M concurrent Sandboxes in its engineering test. A benchmark describes the conditions tested, while a customer's deployment still has its own capacity, startup and reliability requirements. The company's platform description identifies a custom file system, container runtime, scheduler and image builder, with usage-based pricing. That commercial model ties Modal's opportunity to work actually running on its platform. For customers, the evaluation extends from whether the API is pleasant to use to how the platform behaves when many agents need environments at once, with real dependencies and a delivery deadline attached. Model changes bring another operating decision onto Modal. Execution is only one part of Modal's expanding role. Its Endpoint Candidates release, announced October 12, lets teams stage alternative inference deployments and mirror a sample of production traffic to them. Teams can inspect latency and throughput, run their own quality evaluations and promote a candidate while retaining the production endpoint URL. The original deployment continues answering users during evaluation. Modal also explains that mirrored requests are best-effort and may be dropped under pressure, protecting the original production request from a slow candidate. That behavior matters when interpreting results: evaluating a model inside an application involves traffic handling and resource limits as well as the model's output. Together, Sandboxes and Endpoint Candidates put Modal near two recurring decisions: what an agent can run, and which model deployment should serve customers. The commercial attraction is continuity across experimentation and operation, with the same supplier carrying more of the infrastructure work as an application changes. Customers still have to judge cost, isolation, dependencies and service behavior against their own workloads. A team promoting an inference candidate has a very particular decision in front of it: whether the measured change is good enough to become the response its users receive, on the environment Modal is now being paid to keep running. DevCuration Data AI Infrastructure funding, last 30 days. DevCuration's funding database tracked 38 AI Infrastructure rounds totaling $14B in disclosed capital over the past 30 days. Recent deals DevCuration covered: