Introducing ASpec
Subconscious

Subconscious Blog

Research, engineering and product updates on long-running agents.

Wrong by Design: Adversarial Speculative Reasoning Improves Long Horizon Agents

By Hongyin Luo, Wei Fang · · 2 min read

Wrong by Design: Adversarial Speculative Reasoning Improves Long Horizon Agents

ASpec trains the draft model in our speculative decoder to guess wrong on purpose at an agent's decision points, so the target model writes those tokens with the least noise. It raised DeepSWE success by 7.0 points on GLM 5.3 and 4.0 points on DeepSeek V4.1 Flash.

Read post

All posts

Vercel x Subconscious

Subconscious is now on Vercel AI Gateway

Subconscious is live as a provider on Vercel AI Gateway. If you already route through the gateway, you can switch to GLM 5.3 and DeepSeek V4.1 Flash on our inference system with one line of code and get better agent performance.

· Announcements

Product will be built by software factories and secured by exploit factories

Exploit Factories

With an increase in cybersecurity exploits used by cyber attackers, software teams should set up "Exploit Factories" to continuously probe and patch their products before the agent hoards get to them first.

· Developers

Nori x Subconscious Partnership

Nori and Subconscious Partner on Long-Running AI Agents

Nori and Subconscious partner to let teams run long, tool-heavy AI agents in the cloud. Kick off a session in a few clicks and close your laptop.

· Announcements

Funding Announcement

Subconscious Raises $5.1M to Build the Inference Platform for Long-Running Agents

MIT researchers discovered a way to dynamically compress up to 80% of an AI agent’s context to make it much more efficient and accurate. Now, they’re turning that core technology into an opinionated inference platform to power agents that run can faster, for longer, at a lower cost with no changes to the underlying hardware, AI models, or apps that use them.

· Announcements

WATTER + Subconscious

WATTER and Subconscious Power AI Agents with Your Water Heater

We have partnered with WATTER to bring our inference system and models to the WATTER AI platform, making it possible to run capable, long-context agents on a distributed fleet of GPUs built into water heaters.

· Announcements

Subconscious + GLM-5.2 Makes "/compact" Obsolete

Subconscious + GLM-5.2 Makes "/compact" Obsolete

Model-driven context engineering improves agent inference and reasoning

· Agents

Caching engineered context for agent inference

Subconscious Cache: Reliably Capture Your Agent Context

We introduce an inference runtime optimization specifically for agent systems with automated context pruning and the underlying caching technology to make it happen.

· Research

Introducing TIM-Qwen3.6-27B

Introducing TIM-Qwen3.6-27B

Today we're releasing TIM-Qwen3.6-27B on a new OpenAI and Anthropic compatible API, powered by a TIMRUN runtime breakthrough that delivers 10x effective context and 3x concurrent throughput on the same hardware.

· Announcements

A new hope for SLMs

A New Hope for SLMs

Last week Claude Opus 4.7 became the most powerful model ever released to the public. This week it doesn’t seem to matter very much.

· Agents

March benchmarks of our system

Context Intelligence: March 2026 Agent Benchmarks

The benchmarks are in. Training models to manage their own context sets a new frontier across browser control, computer use, and coding.

· Research

trash-man

Fight context rot with garbage collection

Python programs manage memory efficiently with garbage collection. Agents should do the same thing for managing context.

· Agents

dominoes

Why I joined Subconscious

A simple AI agent that helps users click a few buttons shouldn't cost $0.50 per conversation. And building it shouldn't require some of the best engineering work I've done in my 10-year career.

· Announcements

The Model, The Engine, and the Agents

Start Your (Agent) Engines

We think it’s a bit outdated to define an agent only by its underlying model.

· Agents

Better Call TIM

Better Call TIM: Filling out forms with an agent

Thread Inference Model (TIM) outperforms frontier browser agent.

· Agents

The End of the Agent Dev Paradox

The End of the Agent Dev Paradox

We want flexibility and rigidity, we want a do-it-all blackbox but with complete observability, and we want tools for fast prototyping and performance tuning all at the same time in the same agent.

Abstract as High as Possible

Abstract as High as Possible

When you write at a lower level abstraction than you have to, you’re going to regret it.

· Agents

Is GPT-5 always more useful than GPT-2?

Make Agent Frameworks Trainable

Agent frameworks cannot generalize. Something that can is inevitable.

· Agents

Dawn of a New Interface

Dawn of a New Interface

LLM OS is here

· Research

Agents Without Training Wheels: The Model Agent

Agents Without Training Wheels: The Model Agent

It's time to remove the training wheels from Agents. The model agents are here.

· Agents

The Cure for Context Rot

The Cure for Context Rot

Context rot is the slow death of an agent’s attention span, and a key problem holding agents back today.

· Agents

Introducing Subconscious: Build Flexible, Capable Agents

Introducing Subconscious: Build Flexible, Capable Agents

Backed by MIT research on a co-designed AI model and runtime, Subconscious is an inference engine for AI agents that allows for strong reasoning across external tool calls and virtually unlimited context.

· Announcements

Try Subconscious today.

Mint a key, change the base URL, and run your next session on Subconscious.

Start building