Autonomy

Good Voice AI products are hard to build. These 22 lines of code can support 10k+ concurrent voice agents. In Autonomy, agents are modeled as concurrent stateful actors. When you enable voice, the framework creates two actors per agent. A fast voice interface actor listens and speaks. It handles greetings, turn-taking, and filler phrases like "that's a good question." It uses a low-latency realtime model for sub-second interactions. A primary agent actor thinks and acts. It runs tools, retrieves knowledge, and handles complex requests that require multiple autonomous steps. This two-layer design makes an agent feel responsive and a natural participant in conversation. When users interrupt, the voice layer catches it and cancels cleanly. When they ramble or change direction, the primary layer reasons through it. You configure application-level behavior, not pipelines. Tune voice activity detection for natural turn-taking. Write separate instructions for each layer. Add tools to take actions and knowledge sources to ground the agent. Autonomy handles the hard parts: thousands of concurrent audio streams over websockets, audio buffer management and chunk handling, isolated memory per conversation, message ordering so responses never arrive out of sequence, barge-in that cancels cleanly when users interrupt, noise handling when they're calling from a coffee shop or a car. Focus on shipping a great conversational experience, instead of spending months building and scaling complex voice infrastructure. A full step-by-step guide with a live demo, in the comments below 👇

Back in December Michael Hoy 👨🚀, CEO of Atlas, and I sat down to shoot the shit about product pricing. It was a really fun chat...and I finally posted it to YouTube! https://lnkd.in/g7jwPn6p

Gokul is spot on in this post. But the challenge is even bigger. The last gen of vertical AI companies are not just competing against one deep-working long-horizon agent. They are competing against parallel fleets of them. Autonomy enables their competition to create parent agents that can spawn and delegate work to thousands of sub-agents. Each sub-agent has its own filesystem, a shell to run CLI tools, and the ability to write and run new programs on the fly. They divide complex problems, attack from multiple angles, and converge on outcomes in a fraction of the time. Agents, in Autonomy, are modeled as concurrent actors that automatically form secure distributed clusters to enable massive scale on a tiny infra footprint. This creates orders of magnitude advantages in costs, speed, and scale. The question to benchmark is: Can your specialized agent outperform a coordinated team of 100s or 1000s of really-cheap general-purpose agents that can code their way around problems in real-time? If not, then the time to change your approach is now.

Context graphs are the next big AI primitive. "Incumbents can’t build them." Autonomy did! ... This is a sharp piece from Foundation Capital (Jaya Gupta, Ashu Garg) on why context graphs unlock a new breed of application that I call "Autonomous-Native". It’s worth a read if you care about agents that persist and reason over time with the "whys" and "hows", not just sifting though the "whats" that live in Cloud-Native data silos. https://lnkd.in/gFRPYXJf

Gokul is spot on in this post. But the challenge is even bigger. The last gen of vertical AI companies are not just competing against one deep-working long-horizon agent. They are competing against parallel fleets of them. Autonomy enables their competition to create parent agents that can spawn and delegate work to thousands of sub-agents. Each sub-agent has its own filesystem, a shell to run CLI tools, and the ability to write and run new programs on the fly. They divide complex problems, attack from multiple angles, and converge on outcomes in a fraction of the time. Agents, in Autonomy, are modeled as concurrent actors that automatically form secure distributed clusters to enable massive scale on a tiny infra footprint. This creates orders of magnitude advantages in costs, speed, and scale. The question to benchmark is: Can your specialized agent outperform a coordinated team of 100s or 1000s of really-cheap general-purpose agents that can code their way around problems in real-time? If not, then the time to change your approach is now.

So very cool. “The screenshot below is everything you need to give every tenant instance of a multi-tenant agent a dedicated file system in Autonomy .” Great work Mrinal Wadhwa

In 2026, you'll hear from everyone: File Systems enable agents to do deep work. The screenshot below is everything you need to give every tenant instance of a multi-tenant agent a dedicated file system in Autonomy. FilesystemTools(). That's it. The agent can create files and directories, take notes, read a range of lines in a file, grep, glob, and more. 𝗪𝗵𝘆 𝗳𝗶𝗹𝗲 𝘀𝘆𝘀𝘁𝗲𝗺 𝘁𝗼𝗼𝗹𝘀 𝘄𝗼𝗿𝗸 𝘀𝗼 𝘄𝗲𝗹𝗹 𝗳𝗼𝗿 𝗱𝗲𝗲𝗽 𝘄𝗼𝗿𝗸: Agents working on complex, long-running jobs accumulate information fast. Too much in context causes agents to lose focus. They drift off course, miss details, pick the wrong next action. A file system lets agents offload. Write notes to a file. Read only what's needed. Keep the context it gives to its LLM on each turn small and focused. It also gives you precision. RAG and vector search find semantically similar results. But when you need a specific value like side effects of a drug or the address of a customer, `grep` returns exact matches. A file system also allows an agent to preserve natural hierarchies of the data it is dealing with. For example, an agent preparing docs for FDA drug approval can organize by submission, then by section, then by document: /nda-2024/clinical-data/trial-results.pdf. That structure mirrors how the FDA expects submissions. Embeddings would flatten it. Link to more info in the Autonomy docs in the first comment 👇