Bolt Perspective | A Thousand Agents in a Thousand Eyes

When people talk about AI applications, "agent" is probably the word that comes up most often. But whether in research or industry, it means very different things depending on who's using it. Madrona recently published a solid overview of agent infrastructure (link: https://www.madrona.com/the-rise-of-ai-agent-infrastructure/), and there's one line in it that really hits the mark: "Talk to enough practitioners, and you'll find a range of different concepts all being called agents."

That we already have a layered infrastructure stack in a field where the definition itself is still up for grabs probably says something about the expectations we've pinned on it.

If it walks like a duck, it's a duck

Different agent implementations may sit in completely different architectures and carry completely different tasks, ranging from simple to complex. But cutting through the noise, most discussions about agents boil down to two characteristics:

• Autonomous: An agent can perceive its environment, make decisions independently — and that includes reasoning (e.g., CoT, Chain of Thought), reflection (e.g., ReAct), tool use (e.g., Toolformer), and so on. The vast majority of agent research and development work today falls in this bucket. Today's large models don't have these capabilities out of the box; developers need to provide them for specific scenarios.
• Self-improvement: An agent can gradually optimize itself through feedback during operation, such as learning new skills and refining skill combinations. Much of the work in this area is still at the research stage. In practice, agent optimization still mostly relies on human intervention afterward.

The soil that nurtured it is also the shackles that constrain it

Agents stem from our vision of intelligent entities. But today's agents are also largely a workaround for the weaknesses of models themselves — insufficient reasoning capability and limited context windows. At the same time, model reasoning ability is the biggest constraint on what agents can do; their efficiency and success rate on complex tasks remain underwhelming (see the recent examples of AI software engineers). Moreover, today's agents depend heavily on scaffolding built by developers (e.g., CoT) to guide models through tasks. Rather than the model being the brain of the agent, it's more like the pigeon in an early guided missile (during WWII, U.S. military scientists hoped to develop a missile controlled by three pigeons; while there were some successful tests, the pigeon missile proved impractical for combat and the program was eventually canceled).

(Image generated by GPT-4; input from https://en.wikipedia.org/wiki/Project_Pigeon)

Infrastructure vs. agent development

A further leap in reasoning capability isn't on the horizon yet — it requires a breakthrough in model architecture, not model scale. On the timeline for this, we've seen predictions ranging from 2025 to 2030. But when that breakthrough does come, and models can fully grasp decision-making, much of the hand-holding we do today will no longer be necessary. We'll still need a layer of agent infrastructure, but the actual agent development on top of it starts to look like the work of business analysts. Maybe that's closer to what we originally imagined an intelligent agent to be.

Either way, this field is still very early. Both agent infrastructure and agent-based applications will keep emerging. We remain actively engaged with this space. Whether you're a researcher exploring this area, building agent infrastructure, or an entrepreneur applying agents to real problems — welcome to join our community and chat with us (my WeChat: Can_Zheng).

** Linear Bolt Bolt is Linear Capital's dedicated investment program for early-stage, global-market-facing AI applications. It carries Linear Capital's investment philosophy and principles, focusing on projects where technology-driven transformation is at the core. Bolt aims to help founders find the shortest path to their goals — whether in speed of execution or investment approach, its commitment is lighter, faster, and more flexible.