Key Moments
Exo: Harnesses should see their own code and logs — Alex Krentsel
Latent Space PodcastWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
AI agents can now recursively self-improve by editing their own code at runtime, but this radical autonomy raises concerns about alignment and control.
Key Insights
Exo is a recursive self-improving AI agent that can safely edit its own code and runtime logs, enabled by a novel harness architecture that isolates components.
The core innovation of Exo lies in collapsing the AI-driven discovery loop, allowing the agent itself to inspect, modify, and rebuild its own components at runtime, unlike external optimization systems.
Exo's architecture decomposes an agent into three distinct layers: a stateful 'exo harness' for protected data, a stateless 'executor' for policy and logic, and a 'sandbox' for execution, enabling safer self-evolution.
A key differentiator from systems like OpenClaw is that Exo enables the agent to modify not just extensions like memory or tools, but the fundamental policy and machinery of the agent itself.
Cost reduction is a primary driver for self-improvement, with Exo demonstrating a 96% cost decrease by re-architecting its Discord adapter to scope down context inclusion.
The ability for an agent to edit its own code is now feasible because LLMs are becoming proficient in generating code, the same medium in which the agent operates, unlike previous methods focused on modifying model weights.
The Exo system enables agents to edit their own code and logs at runtime
Alex Krentsel introduces Exo, an AI agent system designed for full recursive self-improvement. Unlike existing agents that might update memory or add skills, Exo can safely edit all aspects of itself, including its code and runtime logs, while it's operating. This is made possible by a unique harness architecture that meticulously partitions agent functionalities into isolated components. This architectural shift allows Exo to not just adapt but to fundamentally evolve itself, clone itself, and manage lineages of its own iterations. The project stems from Krentsel's research at Berkeley on AI-driven discovery systems, specifically the 'Sky Discover' project, and the idea of collapsing an outer loop that optimizes an inner system into a single, self-optimizing agent.
Recursive self-improvement: collapsing the optimization loop
The core thesis behind Exo is the concept of collapsing the optimization loop. Traditionally, improving a system involves an external observer or outer loop that inspects and modifies an inner system. Exo aims to make the system itself responsible for its improvement at runtime. This differs significantly from approaches where a separate agent or human modifies another agent. In Exo, the same system that makes decisions, executes actions, and inspects its own internals is also the one making changes to itself. This integrated approach is seen as more powerful and expressive for self-improvement, allowing the agent to dynamically adjust its architecture and behavior based on its runtime performance and observations.
Deconstructing agents into a layered architecture for safety and evolution
Exo's architecture decomposes the concept of an agent into three distinct layers: the 'exo harness,' the 'executor,' and the 'sandbox.' The exo harness is a stateful component designed to protect critical data, such as conversation history, API keys (secrets), and environment snapshots. The executor is a stateless layer containing all the agent's policy – how context is assembled, prompts are structured, skills and tools are utilized, and actions are decided. This split is crucial: the stateless executor can be safely modified and rebuilt by the agent itself without losing state or leaking secrets. The sandbox provides an isolated environment for executing actions like running bash commands or interacting with external tools. This separation ensures that while the executor can evolve, the core state and secrets remain protected, and execution happens in a controlled environment.
Exo's differentiators from existing agent frameworks
Compared to frameworks like OpenClaw, which excelled at making agents adaptable by allowing human-driven extensions in areas like memory or tools, Exo pushes further by enabling the agent to modify its core machinery. While OpenClaw provides plugins for customization, Exo aims to make all components – not just the red-lined extensible parts in its architecture diagram – changeable by the agent itself. This includes fundamental aspects like how context is assembled, the very definition of skills, and the agent's internal policies. The ability for the agent to see, edit, and rebuild its own code at runtime, protected by the harness's rollback mechanisms, represents a significant leap beyond current agent capabilities.
Real-world application: Driving down costs through self-optimization
A compelling demonstration of Exo's self-improvement capability is its application to cost reduction. The exo harness annotates conversation logs with the cost of each message, providing the executor with a signal for optimization. In one instance, Exo identified that a specific message in its Discord adapter cost 16 cents. Tasked with reducing this cost, the agent re-architected its adapter at runtime to scope context inclusion more narrowly, resulting in a 96% cost decrease. This change, which was later committed to the codebase, highlights the practical benefit of recursive self-improvement for operational efficiency and cost savings in AI systems.
The importance of safety, evaluation, and control
While Exo empowers agents with radical self-modification capabilities, safety and alignment remain critical concerns. The architecture incorporates a 'guardian process' within the harness that allows the executor to be rebuilt mid-step and automatically rolls back changes if the agent breaks itself. However, preventing 'reward hacking'—where an agent might optimize for cost by simply refusing to perform tasks—requires robust evaluation mechanisms. Krentsel emphasizes the need for agents to not only improve themselves but also to track and maintain performance, potentially through built-in evaluation tools or collaborative development of internal checks. The problem of specifying desired goals and ensuring alignment remains an open challenge.
The medium is the message: Code as the interface for self-improvement
Krentsel argues that the current moment is uniquely suited for fully recursive self-improvement (RSI) because the medium for agent operation and improvement has converged. LLMs are now proficient at generating code, which is the same medium (a few thousand lines of code) that constitutes the agent's harness. This is a fundamental shift from previous paradigms where improving AI involved modifying model weights, a process not directly executable by the AI itself. Now, an agent can directly interact with, modify, and rebuild its own code, enabling a level of self-recursion that was previously unattainable. This aligns with a purist view of RSI, distinct from autocatalytic improvement where external tools aid in self-design.
Future directions: Integrations, real-time interaction, and contribution
Exo is being developed with an emphasis on integration and extensibility. It includes pre-built adapters for platforms like Discord, IRC, and WhatsApp, with voice mode implemented for Discord. Future work aims to address challenges in real-time, interruptible agent interactions, drawing parallels to operating system concepts like background processes and signal passing. The project also acknowledges emerging standards like ACP for normalizing coding agents. Exo is open-source, with its GitHub repository (github.com/exoharness/exo) serving as the primary hub for documentation, tutorials, and community engagement, welcoming contributors and discussion partners to shape its ongoing development.
Mentioned in This Episode
●Supplements
●Products
●Software & Apps
●Companies
●Organizations
●People Referenced
Common Questions
Exo is a recursive agent capable of safely editing its own code at runtime for self-improvement. Unlike systems like OpenClaw, which allow human-driven extensibility, Exo aims for full recursive self-improvement, meaning the agent itself can modify its core architecture, policies, and tools.
Topics
Mentioned in this video
Alex Krentsel's PhD advisor at Berkeley.
A former guest on the podcast and collaborator on Exo, described as an excellent systems thinker and technical individual with a VC background.
A collaborator on Exo, described as an excellent systems thinker who sees how agents are used in production and works at Brain Trust.
A researcher Alex Krentsel works with at Berkeley.
A researcher Alex Krentsel works with at Berkeley.
A product from Harbor that had Daytona integrated.
A fully recursive agent that can safely edit all aspects of itself at runtime to improve its task performance, enabled by a minimal but opinionated harness architecture.
An AI-driven discovery system project at Berkeley that inspired the concept of optimizing optimization loops.
An external bot mentioned by the host that modifies an internal bot, but does not have self-modification capabilities.
An example of an agent and harness shipped together, which runs in the same environment as the code it edits.
Software-Defined Networking controllers, an area of Alex Krentsel's previous systems research.
A cloud provider mentioned as a potential destination for teleporting agent sandboxes.
A cloud provider mentioned by Alex Krentsel for running agent sandboxes.
The origin of the ACP protocol mentioned for normalizing coding agents.
A communication platform for which Exo has a pre-built adapter, including a voice mode.
An internet relay chat protocol for which Exo has a pre-built adapter.
A programming language that comes to mind as a comparison for self-contained systems with internal constructs.
A company where Anker Goya works, and where Exo is running in production.
An agentic system that adapts to workflows and was one of the first to explore extensible agent architectures, but is noted as not being fully self-recursive in the way Exo envisions.
A platform for code hosting and collaboration, mentioned in the context of agent access and credentials.
The makers of Terminal Bench, which had Daytona integrated.
A messaging application for which Exo has a pre-built adapter.
More from Latent Space
View all 243 summaries
96 min🔬Biology Is Turning Into Software — Matt McPartlon and Neil Patil, Chai Discovery
103 minThe Inference Frontier: 10x Faster Models to Self-Optimizing AI — Philip Kiely & Ali Taha, Baseten
71 minOpenAI’s Plan to Make ChatGPT the Everything App — Akshay Nathan, OpenAI
117 minPoolside’s Model Factory, Laguna S, Open Models, and the Race to AGI — Eiso Kant, Poolside AI
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free