Agentic learning for software engineering
From learning
to engineering
AI can generate the code. OwlHub teaches you how to engineer the outcome.
Accounts are not open yet — OwlHub is in a closed pilot.
$ owlhub review --hunt resilient-ingest-api✓ pass Authentication 30/30✓ pass Authorization 27/30✓ pass Security Config 15/15! review CI & Quality Gates 21/25Grade A 93/100→ 2 issues found. Fix and resubmit.
The problem
Courses teach concepts. OwlHub makes you prove you can build.
A finished course records attendance. It cannot tell an engineering manager whether you can specify a system, direct an agent through it, catch what the agent got wrong, and explain the trade-off you shipped. That gap is what OwlHub measures.
- 18 h 42 m of video watched
- Quiz score 87%
- Final project cloned from the lesson repo
- Certificate of completion issued
- A running service you specified, built and deployed
- Tests you wrote for failure modes the agent missed
- Your reasoning for the trade-offs you shipped
- A written defence of the trade-off you shipped
How it works
One loop, repeated until you've mastered it.
A Hunt turns what you learn into production work. You move through six stages, and each one leaves an artifact behind. Then, every few Hunts, a Checkpoint can stop the loop and ask you to defend what you built.
- 01
Learn what the system demands
Short, dense material tied to the brief in front of you — the failure modes and constraints you will actually hit.
Outputbrief.md · constraints - 02
Build it with agents you direct
Write the spec, plan the work, put coding agents on it. You decide what to accept and what to rewrite yourself.
Outputspec.md · repo · ADRs - 03
Submit the system, not a screenshot
Repository, test suite, pipeline run and deployed environment go in together — not a screenshot of a passing build.
Outputsubmission bundle - 04
AI review against a deterministic rubric
Checks run first: build, tests, performance budget, deploy health. Then the reviewer reads your work against expert-written criteria and tells you which criterion failed, and why.
Outputrubric decision + reason - 05
Improve, and defend what you kept
Fix the failing criterion, or argue it. Then answer for the system you shipped — why this boundary, why this key.
Outputdefence transcript - 06
Advance with the evidence attached
The record travels with you: artifacts, rubric decisions and reasoning, readable by anyone assessing your work.
Outputverified competency record
Every few Hunts, the loop can stop and ask you to defend the work.
OwlHub is being designed around system-aware checkpoints that test whether you can explain, debug, and adapt what you built — not simply submit working code.
Finishing Hunts gets you to a Checkpoint. The Checkpoint is where you show the work is yours.
“Your ingest service drops a downstream region. Walk me through what happens to in-flight events — and what you would change.”
from hunt/resilient-ingest-api“Your handler dedupes at the store, not in memory. What breaks if you move it?”
from ingest/handler.ts“That tenant doubles overnight. What breaks first, and how would you know before your users do?”
from adr-004 · partition strategy
Answered in writing, against the same rubric standard as the Hunts. Illustrative questions.
Why it is different
Designed to measure engineering, not attendance.
Real production systems
Hunts carry production constraints: latency budgets, failure modes, migrations, rollbacks, cost. No toy to-do apps, no sandbox that forgives what production would not.
Deterministic rubrics
Criteria are written by engineers before you start, and applied the same way every time. Machine checks are separated from AI-assisted judgment, and a failing criterion is never averaged out of sight.
AI-native workflow, human accountable
You are expected to use agents — and expected to catch them. Specify, plan, direct, review, and own the result. The skill being trained is direction and review, not typing speed.
Proof over certificates
What you leave with is an evidence record — artifacts, rubric decisions, reasoning — that a reviewer can inspect. A certificate asserts. Evidence demonstrates.
Start your first Hunt.
Pick a track. Build something real. Get reviewed. Master.
Accounts are not open yet — OwlHub is in a closed pilot.