You’ve probably felt it. You ask your AI coding agent—Codex, Claude Code, whatever—for a tiny tweak to the plan. Just one small adjustment. You hit enter, expecting a quick update, and instead, the agent hands you a completely rewritten manifesto.
You stare at the screen. You’re too tired to read the whole thing again. You just approve it, hoping for the best, and immediately lose whatever trust you had in the process.
We thought AI would eliminate grunt work, but instead, it just created a new kind of exhaustion: the fatigue of babysitting a machine that never stops rewriting its own homework.
This is the dirty secret of the AI coding revolution. The agents are brilliant, but they suffer from severe plan drift. They lack context stability. The moment you introduce a minor change, they rip up the blueprint and start from scratch. You’re no longer the engineer; you’re the exhausted proofreader of a chaotic intern.
When you ask for a one-line code change and get a 500-word plan rewrite back, you haven’t gained an assistant—you’ve gained a liability.
This is why a new open-source tool called BDFL is so interesting. BDFL stands for Benevolent Dictator For Life. It’s a terminal supervisor for Codex, Claude Code, and Ollama. And the name is entirely intentional.
The open-source world usually champions decentralized, collaborative, democratic workflows. But when it comes to managing AI agents, that ethos is a death sentence. You can’t collaborate with a machine that constantly changes its mind. You need strict oversight. You need a dictator.
The only way to manage the chaos of autonomous AI isn’t more collaboration. It’s a benevolent dictatorship.
BDFL solves the plan drift problem by doing something painfully obvious yet entirely missing from current AI tools: it versions the plans. You work with a planning agent, and when you request a change, you can review exactly what changed. You can approve individual sections. You can lock sessions so they don’t mutate behind your back.
Once approved, BDFL spins up isolated worker agents to implement the work. The supervisor handles the scheduling, the checks, the integration, and the recovery if things break. You get the flexible, iterative changes you want, but packaged in stable, reviewable increments.
There is a tension here, of course. By introducing a layer of granular approval and isolated execution, BDFL risks adding overhead. It might slow down the very speed that makes AI agents so intoxicating in the first place. But let’s be honest: speed without control is just crashing into a wall faster.
If you’ve been burning out on opaque AI plan rewrites, BDFL offers a lifeline. It restores human control over the planning process. It forces the AI to work within your boundaries, not the other way around.
Autonomy without oversight isn’t intelligence. It’s just chaos with a UI.
The era of blindly trusting AI agents to figure it out is over. If we want to actually ship code with these tools, we need to stop treating them like peers and start treating them like what they are: powerful, unpredictable engines that require a firm hand on the wheel.
FAQ
Q: Isn't adding a supervisor just going to slow down my workflow?
A: It adds a necessary layer of friction. Yes, it might slow down the initial execution, but it saves you the hours of debugging and re-reading massive plan rewrites. Speed without control is just crashing faster.
Q: What's the practical implication of using BDFL?
A: You regain visibility. Instead of blindly approving massive AI-generated plans, you can review specific diffs, lock sections, and let worker agents execute in isolated environments. You stay in control of the architecture.
Q: Is the 'Benevolent Dictator' naming just a gimmick?
A: No, it's the core philosophy. Decentralized, collaborative workflows fail with AI because machines don't have intuition. Managing AI plan drift requires strict, centralized, authoritarian oversight to force the AI to stick to the approved blueprint.