Agentic AI Frontiers: Long-Horizon Agents, Voice Agents, and Benchmarking Breakthroughs

Long Horizon Agents Build Day

The frontier of AI is no longer measured in tokens, it's measured in time.

The length of tasks an agent can complete autonomously is doubling roughly every seven months. What ran for seconds a year ago now runs for hours. The next leap is agents that pursue goals across long horizons — holding context, self-correcting, and staying coherent across an entire workday.

Building agents that stay reliable over long horizons is one of the hardest open problems in AI. That's what we're here to build.

In May, we teamed up with Coframe for the Internet of Agents Build Day — a packed day of builders shipping the agentic web. Now we're back to go deeper: AGI House is bringing together the best builders for a one-day sprint to push the boundary of what agents can do over long time horizons.

Voice Agent Build Day Recap

The next consumer interface may be voice. But making an agent speak naturally is no longer the hardest part.

At AGI House’s Voice Agent Build Day, more than 100 founders, engineers, and researchers explored the product and infrastructure questions that emerge when voice agents enter the real world: latency, interruption, evaluation, customization, cost, and scale.

Kylan Gibbs, CEO and Co-founder of Inworld, shared why consumer voice adoption is accelerating and what builders need to understand about latency, quality, and retention.

Research Spotlight Begins with a Harder Question for Coding Agents

SWE-bench helped the field measure whether AI agents could work inside real codebases and resolve software issues. But what comes next?

In the first episode of AGI House’s Research Spotlight, Jessica Chen sat down with Meta AI researcher Kilian Lieret to discuss why the next generation of coding-agent benchmarks may need to move beyond clearly defined tasks and fixed answers.

Their conversation covers SWE-agent, mini-SWE-agent, CodeClash, and ProgramBench. These projects explore a more open-ended question: can agents pursue a goal, interpret feedback, understand why an approach failed, and improve over multiple rounds?

Research Spotlight is a new AGI House series bringing frontier researchers into practical conversations about the ideas and systems shaping AI.

Before Discovery Loop: Revisiting a Dinner with Oriol Vinyals

Months before co-founding Discovery Loop, Oriol Vinyals joined an AGI House dinner on world models, agents, and the path to AGI.

As Discovery Loop sets out to automate experimentation in science and engineering, we’re revisiting a conversation about how agents learn from experience, close feedback loops, and accelerate research.

 AGI House Talent Board

We’re starting a talent campaign to help high-signal builders discover roles across our portfolio and extended network.

An AGI House Ventures portfolio startup is hiring:

TenX Semi — building AI-powered agents and tools for chip design.
Open roles: AI Systems Engineer, AI Application Engineer, Head of AI
Strong fit for engineers with experience in LLMs, AI agents, code generation, RAG, document understanding, RTL, verification, or chip design workflows.

Stay Updated with Us