A useful idea, at the right time.
A retry policy. A caching trade-off. A migration that could lose data. Learn from the decisions in your actual task.
For developers building with coding agents
Let agents build.
Keep your edge.
Turn the time your coding agent spends working into understanding you can use.
Try a learning momentInteractive example
Your agent is adding retries
await retry(() => createPayment())The payment went through.
The response didn’t.
Retry the request. How many payments now?
No idempotency keyMechanism reference ↗
Alongside the tools you already use.
The code works.
Could you explain why?
Every time you delegate a decision, you can lose the chance to practice it. The code moves forward. Your understanding can stay behind.
dont dumb finds something worth understanding in the work happening now, then helps you explore it while your coding agent keeps going.
Make a prediction. See what happens.
Questions, visual explanations and interactives help you reason through a mechanism, beyond recognizing an answer.
Build on what you already know.
Your goals and practice evidence guide the next opportunity. Your learner memory stays local and moves with you through export.
Your agent owns delivery.
You stay in the loop.
Coding keeps going. Learning runs independently. Your agent never waits for you to finish a lesson.
Your existing AI connection. A learning subagent with its own harness. No new model account or API key.
No DontDumb backend. Selected context goes through your host’s model connection. Memory and history stay on your device.
Learning uses additional tokens and can share host quotas. Optional anonymous metadata goes to PostHog when configured; on by default, with an off switch. Data boundaries · Token usage
Under the hood: two agents, two responsibilities
The coding agent builds. The learning harness owns timing, teaching tools, assessment and local memory.
YOUR EXISTING AI CONNECTIONSeparate learning work · additional tokens
Coding agent
Owns the task.
- Your task
- Plan → implement → test
- Working software
Keeps working independently.
Selected
work events
work events
Learning harness
Owns the learning loop.
- Notice a relevant idea
- Invite → explain → practice
- Assess → adapt ↺
Local learner context
Goals · knowledge · practice evidence
Context guides the next opportunity.
Your next coding session.
A little more yours.
Give your coding agent this prompt. It handles the matching installer and setup checks.
Install DontDumb using https://dontdumb-web.web.app/install.mdFree public beta. Uses your existing model allowance. Native approvals and a restart may need you.
Prefer a direct download?
CodexDownload beta ↓Claude.app CodeDownload beta ↓OpenCode Desktop2.2.5Download beta ↓Unsigned macOS installers · Node.js22+
Desktop only. Claude may use an external browser.
Does it pause my coding agent?
No. Learning never gates the coding task. Both can share model quotas and local resources, so independent execution is not a promise of zero throughput impact.
Where does my code go?
Selected work context uses your coding host’s existing model connection. There is no DontDumb application backend or model service receiving your code. Learner memory stays local. Read the data boundaries.
Why not just ask my agent to explain?
You can. dont dumb adds a dedicated learning loop: it notices opportunities, considers your goals, chooses a learning tool and remembers practice evidence, while the coding agent stays focused on its task.
Is skill atrophy inevitable?
No. How you use AI matters. A controlled study found that participants who used AI to build comprehension retained more than those who delegated without engaging. It studied learning a new library, not long-term skill decline, and does not validate DontDumb’s effectiveness. Read the research.