The honeymoon phase of agentic coding tools like Claude Code and Codex often ends in cognitive overload. As developers delegate more tasks to AI agents, the sheer volume of asynchronous threads can fragment human attention, turning the terminal into a chaotic juggling act. Sergey Kuhne, a developer grappling with this mental exhaustion, has proposed a structural solution that treats context management as a data problem rather than a mental endurance test.
The Problem With Parallel Agents
Kuhne describes a common workflow failure: starting a task, waiting for the AI to reply, and starting another to fill the gap. When replies arrive with requests for clarification or side tasks, each demands its own tab. Developers find themselves hopping between 20 active threads, trying to keep multiple contexts in their brain RAM. The result is exhausted focus and a loss of strategic direction, where the human forgets the original goal amidst the noise of micro-decisions.
Git as the Single Source of Truth
To combat this, Kuhne reorganized his work into a priority queue backed by Git. Every ticket is a markdown file committed directly into the repository next to the relevant code. This ensures that if a session dies, the state of the board remains intact. An agent cycles through this board, researching new tickets and building those tagged as safe by a human. The agent runs tests and browser checks but never deploys to production, reserving that final judgment for the developer.
The Owner Inbox and Artifacts
The system separates execution from decision-making. Anything requiring human judgment lands in an owner inbox, a prioritized list generated by a script reading the entire board. The interface for this inbox is a Claude artifactβa single page rebuilt from Git on every cycle. This allows the developer to open one link, review the change and its rationale, and issue a verdict without scrolling through terminal logs. The verdict is then committed back to Git, closing the loop.
Why This Workflow Matters
The core innovation here is the enforced closure of context. By picking one item, loading its specific context, making a decision, and then returning it to the queue, the developer prevents tickets from hanging in memory. While Kuhne admits the system is still exhausting and faces challenges with tooling maturity and blocked tickets, it restores a sense of focus. It shifts the human role from active participant in every thread to a high-level reviewer of a structured pipeline.
Key Takeaways
- Persistence is Critical: Storing state in Git rather than terminal tabs ensures no context is lost during session crashes.
- Separate Execution from Judgment: Agents should handle safe builds and tests, while humans retain control over deployment and architectural decisions.
- Unified Review Surface: Using generated artifacts to consolidate review needs reduces the cognitive load of switching between multiple chat windows.
- Context Closure: Explicitly closing tickets by committing decisions back to the queue prevents mental clutter and improves focus.
The Bottom Line
We need to stop treating AI agents like chatbots and start treating them like junior engineers with strict ticket systems. Kuhne's approach proves that better infrastructure, not just better prompts, is the key to sustainable agentic development.