Coding agents make engineers faster and, quietly, worse at the job. aiworklab is the workspace that closes tickets and closes the skill gap: one beat of friction exactly where it teaches, none where it doesn't, on top of the agents you already run.
Private beta now · Public beta Q3 2026 · GA early Q4 2026
aiworklab is a teaching layer on top of the agent harnesses you already trust.
Every AI coding tool scores itself on lines accepted and time-to-first-diff. Nobody is scoring whether the human on the other side of the tab key is getting better. In our data, the two are slightly anti-correlated.
Every claim on this page links to the report it comes from. Eight papers, published data, limits sections first.
Pick a mode per task. The agent stays the same; your relationship with it changes. Concepts you have already demonstrated pass through silently. New ones get exactly one beat of friction, at the moment it will stick.
Standard agent loop. Plans, executes, edits. The agent works, you ship. Concepts you encounter are silently logged to your skill graph for later review. No interruptions.
Same speed, with one beat of friction. Before applying any non-trivial diff, a 15-second comprehension check tied to that exact change. Pass and merge. Most learning happens here.
The agent withholds. You write the code; it reviews, points to bugs, asks Socratic questions, refuses to fix things for you. Slowest mode, deepest learning.
For agent-authored changes above a novelty threshold, the merge button stays disabled until you write a 2 to 3 sentence explanation of what the diff does and why. An LLM judges it against the diff.
Active recall is one of the most-studied learning techniques in cognitive science. Tab-to-accept skips it entirely. We put it back.
Concepts already marked demonstrated on your skill graph never trigger a check. The friction shrinks as you grow.
When you genuinely don't have time, you skip. We log it. Your weekly retention report shows the trade-offs honestly.
Product preview with sample data. The check appears only for concepts you haven't yet demonstrated.
Sign in with the LLM provider you already pay for, or point us at a local model running on your laptop. We never proxy or mark up inference traffic. This is a permanent design choice, not a launch limitation.
Paste your API key. Validated client-side and stored in the OS keychain. Nothing transits our servers.
One click. We never see your raw credentials. Tokens stay on-device. The easiest path for non-power users.
Auto-detects models running on localhost. Works fully offline. The only tier where regulated industries can ship.
Your admin provides a gateway URL; the team signs in via SSO. Standard for Team and Enterprise tiers.
An anonymised, aggregated dashboard of skill coverage across your engineering organisation. Concept retention curves. Bus-factor warnings on knowledge held by 2 or fewer engineers. The artefact that closes the budget conversation.
Engineers and leaders who ran aiworklab on real backlogs, in their own words.
We gave aiworklab to two teams running a head-to-head experiment. After eight weeks the coached group had higher solo throughput and better retention numbers. The other group closed more tickets. Leadership had been picking the wrong metric for a decade.
"I've shipped TypeScript for six years. aiworklab surfaced a gap in how I thought the compiler resolves conditional types that I didn't know was there. That's the product working as designed, and it's uncomfortable in exactly the right way."
"Coach mode on a Saturday morning is the closest I've come to real deep work since Copilot became the default. The agent withholding is a feature, not a bug."
"The explain-to-merge gate caught a race condition in a diff my agent produced that I had just accepted without thinking. That's worth the subscription on its own."
"The skill graph surfaced three concepts I'd been avoiding for months by leaning on the agent. Seeing them labelled 'encountered, not demonstrated' was humbling. Two of them are 'mastered' now."
Quotes from private beta participants, shared with permission and lightly edited for length. Roles and locations are as self-described; surnames are abbreviated for privacy and portraits are illustrative stock photography, not the participants themselves.
Join the engineers and teams who think the next phase of AI tooling should be measured by what it does to the human, not just for them. Seats are released weekly.
No commitment · a person writes back, usually within a day
