# AI Makes Us WAY Faster but Fries Our Focus (Here's What I Do About It)

## Executive summary

The video addresses the 'fried brain effect'—the cognitive exhaustion caused by excessive context switching when utilizing multiple parallel AI agents. While AI dramatically increases productivity, the current workflow requires constant monitoring and input, leading to diminished focus. The proposed solution is shifting focus from running more parallel agents to building larger, more autonomous 'harnesses' (like the speaker's Software Factory). These advanced systems are designed to allow agents to run for extended periods, batching decisions and requiring less constant human babysitting, thereby solving the core problem of attention fragmentation.

## Key takeaways

- The Problem: Context Switching Fatigue: The primary issue with current AI usage is not the lack of productivity, but the cognitive load from constantly switching attention between multiple parallel agent sessions. This constant context switching is described as 'killing our attention spans' and leading to exhaustion.
- The Limitation of Current Agents: Current LLMs and agent harnesses are not yet capable of running autonomously for full days. Users are forced to 'babysit' and provide input constantly, which necessitates running multiple sessions in parallel to maintain perceived productivity.
- The Solution: Autonomous Harnesses: To mitigate context switching, the focus must shift to building larger, more robust harnesses that allow agents to plan, implement, and validate tasks autonomously over long periods. This requires investing time in the larger system architecture rather than simply adding more parallel sessions.
- Future State: Primary Session Focus: The ideal future state involves a 'primary session' for high-level planning and validation, while background agents handle the heavy lifting. This is predicted to be aided by LLMs offering near-instant responses (hundreds of tokens per second).

## Technical details

- Agent Autonomy and Harnessing: The speaker advocates for moving beyond simple agent sessions by developing a 'Software Factory'—a custom, larger harness. This factory is designed to codify the user's specific workflow, allowing the agent to manage the entire loop (plan, implement, validate) autonomously, minimizing the need for manual intervention.
- Limitations of Simple Commands: The speaker notes that simple built-in functions, such as `/goal` (e.g., in Codex), are insufficient because they lack the necessary customization to account for unique user validation processes and complex business logic.
- API Integration for Guardrails: For financial control, the speaker implemented a system where the 'second brain' interacts with the Revolut Business API. This method uses physical card limits (e.g., $50 weekly spend limit) and webhooks to enforce spending guardrails, which is more secure than relying solely on prompt instructions.

## Practical implications

- Instead of maximizing short-term output by running more parallel agent sessions, dedicate time to improving the core 'harness' or workflow logic.
- Focus on building systems that allow agents to operate in a closed loop (plan -> implement -> validate) for extended periods, thereby batching human decision points.
- For financial or critical processes, implement physical or API-level guardrails (like spending limits) rather than relying solely on prompt instructions.

## Topics

AI Productivity, Context Switching, Agentic Workflows, Software Architecture, Cognitive Load, Dynamous Agentic Coding Course 2.0, AI Software Factory (open source), Revolut Business

Source: https://www.youtube.com/watch?v=zj1TJlP7c9Y