I need some advice in building a harness

Hey guys, One of the biggest problems I’ve noticed with autonomous agentic development is that execution has become much faster than human understanding can keep up with. For example, I had an agent open 50+ PRs overnight while I was asleep. By the time I woke up, reviewing them became a nightmare—not necessarily because the code was bad, but because I had lost the context behind why certain decisions were made, how one change led to another, and what the agent had learned along the way. That’s the problem I want to solve: keeping autonomous development continuously in sync with the developer’s understanding. The idea is to break work into individual task cards where I can define the objective, choose the agent, set guardrails, token limits, repository scope, permissions, and other constraints. The agent can then execute autonomously, but instead of only producing code or a PR, it also maintains a structured record of its work—a decision tree, execution summary, key decisions, assumptions, and a timeline/MoM of what happened during the task. Over time, every task and session would become part of a persistent graph. Starting from Day 1, I should be able to navigate through the history of the project and understand: What was changed? Why was it changed? What previous decision led here? What did the agent know at the time? And what should I know before approving the next step? Essentially, I want the developer’s mental model of the codebase to evolve alongside the agents working on it, rather than having the agents race ahead while the human is left reconstructing context afterward. I’d also love inputs on what else could be built here to further reduce the gap between agent execution and human context—especially as autonomous agents start running for hours or days without constant supervision. Repo - github.com/kedarvartak/gorilla