aiworklab wraps the agent harness you already use and overlays a teaching layer that nobody else builds. Here's how every piece works, and how they fit together.
The agent is the same: Claude Code, Codex, T3 Code, OpenCode. Your relationship with it changes. Default is Copilot; the system also auto-suggests modes based on the novelty of the files you're touching.
Standard agent loop: plans, executes, edits, runs commands. No interruptions. Concept exposure is silently logged to your skill graph for later spaced retrieval.
Same speed as Autopilot, with one 15-second comprehension check before applying any non-trivial diff containing concepts you haven't yet demonstrated. Pass and merge.
The agent withholds. Reviews your code, points to bugs, asks Socratic questions. Refuses to fix things for you. Heaviest skill-graph updates, deepest learning per minute.
Not LeetCode topics. A per-user, per-repo graph of programming concepts that have appeared in your real code. Each node has a state machine that's honest with you.
The concept appeared in code you accepted but didn't engage with. Logged silently from Autopilot mode.
You read or were given an inline explanation card when the concept came up.
You passed a comprehension check on it, or you wrote it yourself in Coach mode.
You demonstrated it across multiple spaced retrievals over 30+ days. Mastered concepts never trigger checks.
For agent-authored diffs above a configurable threshold (lines changed, files touched, or novelty against your skill graph), the merge button is gated. You write 2 to 3 sentences explaining what the diff does and why. An LLM judges the explanation against the diff.
The concept advances toward "demonstrated." Merge proceeds. Total cost: about 30 seconds.
You see what the judge thought you missed. You can revise, or escalate to Coach mode for that hunk.
You can always skip. We log it. Your weekly retention report shows the trade-offs honestly. No nags, no shame.
Product preview with sample data.
An FSRS-based scheduler picks 3 to 7 concepts due for retrieval each day. Prompts are extracted from code you committed weeks ago, not synthetic exercises. The result: review feels like reviewing your own work.
The modern, open scheduling algorithm that Anki adopted. Replaces SM-2 and calibrates well from a handful of reviews.
"Two weeks ago you wrote this query. Without looking, why did you choose a window function over a self-join?"
An optional weekly 30 to 60 minute window where the agent is read-only. Solo throughput is your headline metric, tracked over time.
Two weeks ago you wrote a request handler that aborts on disconnect. Without looking, what's the difference between the cancellation token and the abort signal you used?
In user_repo.ts you prepared statements once at module load. What's the failure mode if the connection drops?
Recall: in a Yjs document, what happens when two clients edit the same line offline and reconnect?
An anonymised, aggregated dashboard of skill coverage across your engineering organisation. The metric is concept-level, never source code. Available on Team and Enterprise tiers.
The gap is the cohort effect of unmetered AI assistance. aiworklab closes it without sacrificing throughput.
We don't rebuild the agent loop. We integrate with the agent harnesses you already trust, contribute upstream where the licence allows, and put our work in the layer above.
Anthropic's agent loop, integrated through the official Claude Agent SDK. Our default backend at launch.
Codex's open-source CLI agent, integrated via a thin adapter. Reasoning levels and supervised mode supported.
Theo's open-source GUI for agentic coding. We layer on top of its session model.
The open-source agent harness. Full integration including its provider abstraction.
Adapter model. Each harness implements a small interface: start session, plan, tool call, diff, finish, pause, resume, inject. The teaching kernel hooks the diff event. This isolates harness churn from our product. These harnesses are independent third-party projects; we are not affiliated with or endorsed by their developers.
Private beta is in flight. Public beta opens in Q3 2026, and general availability lands by early Q4 2026. Drop us your email and we'll write back personally.
