Tokui Labs Coming soon · pilots opening

// agentic engineering assessment

Measure agentic engineering skill on a real AI product.

Candidates get a running agentic product (an agent with tools, retrieval, evals and traces) and its production issues: wrong answers, broken tool calls, evals that pass while users complain. With a coding agent at their side, they debug, fix and ship. You see how they find the root cause, what they verify, and how they recover.

For engineering teams that hire. We'll write once, when pilots open.

agentic-skills-report.md

Timeline

00:02INC-1 · support agent cites a retired policy
00:06Opened traces · retrieval returns the stale doc
00:11Coding agent's fix compiles · evals green
00:14Replayed real user queries · still wrong 1 in 5
00:22Fixed the index · added an eval that fails first
00:26INC-1 shipped → production passed
00:27INC-2 · agent loops on refund tool calls
Root-causingStrong Eval designStrong Directing the coding agentMeets RecoveryMeets
prompt 3 · "evals are green but they never query the 2025 policy. add one that does"
01

A real agentic product

Not a toy repo, not puzzles. A running AI product with an agent, tools, retrieval, evals and traces, spun up fresh for every candidate.

02

Real production issues

Hallucinated answers, stale retrieval, tool loops, evals that cannot fail. Several incidents, the kind your team gets paged for.

03

Evidence per skill

Root-causing, eval design, directing the coding agent, recovery: each gets a level, with prompts verbatim and what production said. You decide.

Puzzles measure what the agent already does. We measure what the engineer adds.