Introducing Subconscious: Build Flexible, Capable Agents

Inference Designed for Agents

AI agents have yet to deliver. They’re complex to build, inefficient and costly to run, and fragile because of brittle workflows and poor performance on multi-step tasks. Engineers are banging their heads against the wall to deliver agents that were promised but are beyond what’s possible.

That’s why we’re excited to share Subconscious, our platform for running agents with external tool use and long horizon reasoning beyond existing context limits. Backed by MIT research on a co-designed model and runtime, our inference engine decomposes complex tasks, uses external tools without a round trip to an external server, and intelligently prunes its own context to allow for long reasoning chains. Where most teams have built agents by piecing together language model APIs and frameworks at the app layer, our team innovated at the model and runtime layer to solve the problems holding agents back at their source.

This efficient and capable system allows for a much more straightforward developer experience: kick off an agent with a single API call. No more unnecessarily complex multiagent systems. No more context engineering headaches. You define your goals and tools, Subconscious handles the rest.

We’re releasing a research preview of our platform today with a playground and an OpenAI compatible API, and we’d love for you to try it yourself and share your feedback.

A New Generation of AI Agents

Let’s back up for a second. We define an agent as:

A system that takes in goals and tools, then gets the job done with minimal to zero human intervention.

It should be as simple as that. Our agent takes a user request (goals and tools), generates the workflow on the fly, reasons through edge cases, and gets the job done. This is fundamentally different from prevailing architectures, and requires a shift where you as the builder to focus more on the tools you’re giving the agent and also allow the agent more control over its own workflow. We recognize this may be uncomfortable at first, but we also think it's inevitable to solve the world's messiest problems. It allows for a system that's more efficient, more capable, more flexible, and simpler to build on.

Our inference engine is powered by two core components that make this experience possible:

• TIM, our family of fine-tuned models and an acronym for the Thread Inference Model, handles task decomposition, long horizon reasoning, and tool use. No multi-agent orchestration, no manual context juggling.
• TIMRUN, the Thread Inference Model RUNtime, efficiently reclaims memory at scale and delivers low-latency tool execution across threads.

Together, these components allow you to run long-context, tool-using agents as a single-agent system that is efficient, flexible, capable, and fast.

With Subconscious, teams can stop wasting time extending agent frameworks and start focusing on what makes them unique. We are just getting started, but our platform is already working for companies today.

• Stack AI is listing Subconscious as a provider to offer simpler automation to their end users.
• Brix is using Subconscious to search and understand millions of resumes, connecting candidates to roles with agents that can process rich, long-form data.

We think every vertical AI company, from sales to cybersecurity to accounting, will need our infrastructure to compete.

Start Building Today

Today we’re sharing two versions of our inference engine, TIM-small and TIM-large, both available behind OpenAI compatible APIs.

• TIM-small-preview is our core fine-tuned TIM-8b model running on TIMRUN. It enables long horizon reasoning and tool use, and is tuned specifically for tasks with search tools.

• TIM-large is a highly capable generalized inference engine built on the same principles as TIM-small but backed by the power of OpenAI GPT-4.1. We talked to teams around the world building agents, and many won’t deviate from using anything but the most powerful generalized models. We listened, and we built a way to experience long horizon reasoning and the agent developer experience we imagined backed by the power of models you’re more familiar with.

Teams can experiment with TIM, integrate it into their existing workflows as a single capable node, or use it as the core engine for their agents. We’ve had great results experimenting with TIM on search, research, browser use, and other agentic tasks, and we’re excited to see what you build.

In addition to our platform and API, there are two ways to learn more about our research and our platform:

• Technical Report — Deeper technical dive and benchmarks on the system’s performance
• Docs - A set of docs and examples to get you started with our API

This is the very beginning, and we’re excited to pave a new path to build efficient, flexible, capable and simple agents. If you’re serious about building agents, we’d love to talk.

- Jack and Hongyin

(Book a quick chat with us here if you’re building agents)