A "React for Agents" and the Self-Improving Harness

Arjun began his demo by polling the builders in the room, asking how many were running an agent product in production. "That's lesser than expected," he remarked. This poll served as the entry point to his core prediction: the way agents are built is about to fundamentally change.

From Manual Tuning to Meta-Artistry

The common loop for agent development involves running the agent, evaluating its performance, and manually tuning its logic or prompts. Arjun’s talk suggested this entire cycle is on the verge of being automated away. His prediction is that builders will stop writing the fine-grained logic for agents directly. Instead, agents will alter their own harnesses and self-improve based on real-world interaction.

He pointed to an emerging idea called the "meta-artist." In this new role, the builder’s job is not to perfect a single agent's performance through manual effort. The task becomes designing the environment and the initial structure that allows an agent to improve itself.

This requires a different kind of architecture, one that isn't a bottleneck to the model’s own capacity for improvement. I've noticed the focus of our sharpest builders shifting from what an agent can do to what it can become, and Arjun's model for self-evolving harnesses gives that shift a concrete architecture.

A "React for Agents"

To enable self-improvement, Arjun argued for a harness that is “very modular, composable, and very easy to alter.” He drew a direct analogy to web development. “It's almost like React,” he explained. React succeeded because it made it easy to alter UIs by composing different components.

Arjun thinks of an agent’s harness as its interface to the world. Just as interacting with a webpage changes the state of an application, an agent's actions change the state of the world it operates in. He gave the example of shopping on Amazon. When you add an item to your cart or complete a purchase, you are making a change to the state of Amazon's warehouse, which then ships the item to you. The harness is what allows the agent to make these state changes.

This thinking leads to a component-based model for the agent itself. During his demo, Arjun showed a template agent assembled from distinct, pluggable components: one for compacting the context window, one for managing budgets like token limits, one for writing and executing code, and one for communicating with other sub-agents. Any of these pieces could be swapped out or modified, making the agent’s structure flexible.

Evolutionary Search Over Harnesses

With a modular harness, an agent's architecture becomes a variable that can be optimized automatically. Arjun demonstrated this with an evolutionary algorithm. The process begins with a base agent harness. An orchestrator agent then runs that agent multiple times, each time with different variations of its components.

After these runs, the orchestrator inspects the outputs, identifies the best-performing variation based on "any metric that you're trying to improve," and uses that superior candidate as the starting point for the next cycle. "As you just keep doing it in a cycle,"

Arjun explained, "you can kind of get an optimized harness." This creates an automated loop where the system actively explores different configurations to find more effective designs, without a human manually tuning every parameter.

Trading Off Conflicting Goals

This automated approach introduces its own challenges. An audience member raised a critical point about conflicting objectives. For example, a builder might want to optimize for the quality of the agent's output while also optimizing for token efficiency. These two goals can work against each other. better research and reasoning often require more tokens, which increases cost.

Arjun’s response clarified that the goal is not to find a single "best" solution but to map the landscape of tradeoffs. He explained that multi-objective genetic algorithms can be used to find the Pareto frontier,the set of optimal solutions where improving one metric necessarily means degrading another. The builder is presented with a menu of these efficient tradeoffs, not a single winner. You can then choose the point on that frontier that best fits your specific needs, whether prioritizing cost, quality, or a balance of the two. A composable harness makes it possible for an automated process to explore this complex landscape.

If agents can begin to optimize their own internal structures, the role of the agent builder

transforms. The work shifts away from the tedious, manual tuning of prompts and logic and toward designing the evolutionary environment itself. The key question is no longer "Is this prompt working?" but rather, "Is this harness modular enough for the system to explore?" and "Have I defined the right fitness criteria for the agent to navigate?" When the harness is no longer a static constraint but a dynamic part of the solution, agents do not just execute tasks, but actively improve how they perform them.

---

SQ Collective hosts Coworking Fridays for founders, operators, and AI builders working through real product questions in Singapore.

Join an upcoming Coworking Friday: https://lu.ma/ai-labs

Explore SQ Collective: https://www.sq-collective.com

Michael