The Hypothesis

Can an AI agent start as a junior team member, learn from a human expert through questions, and grow into full autonomy?

The Concept

Inspired by the smol-ai/developer approach — what if instead of giving an agent full capabilities from day one, you started it as a junior staff member? A junior email marketer, for example — one that knows the basics but works alongside a human expert, asks questions when it's unsure, and gradually builds competence through real work. The agent learns not from training data, but from mentorship.

Junior agent apprentice

The hypothesis

Can an AI agent start as a junior team member, learn from a human expert through questions, and grow into full autonomy?

The concept

Inspired by the smol-ai/developer approach — what if instead of giving an agent full capabilities from day one, you started it as a junior staff member? A junior email marketer, for example — one that knows the basics but works alongside a human expert, asks questions when it’s unsure, and gradually builds competence through real work. The agent learns not from training data, but from mentorship.

How it works

1. Assign junior role — agent starts with limited scope (e.g. draft subject lines)
2. Agent works and flags uncertainty — completes tasks, asks the human when unsure
3. Human reviews and answers — expert corrects, explains, sets guardrails
4. Agent absorbs context — builds a working memory of preferences, patterns, decisions
5. Expand scope gradually — agent takes on more as confidence and accuracy grow

A mentorship model — the agent earns autonomy through demonstrated competence, not configuration.

What it explores

• Can an agent meaningfully learn from a human mentor in real time, not just from pre-training?
• What’s the right interface for an agent to ask questions without being annoying?
• How fast can a junior agent reach the competence of a fully-prompted one?
• Does the apprenticeship model produce better long-term performance than giving full autonomy upfront?
• What does the human expert need to see to trust the agent enough to let go?

What we found

• Apprenticeship agents made fewer critical errors than fully-prompted ones — the questioning phase caught edge cases a system prompt would miss
• Ramp-up period was 2–3x longer than giving the agent full context upfront
• The agent’s questions were often more revealing than its answers — uncertainties highlighted gaps in the brief the human expert hadn’t noticed

Learnings

• Agent uncertainty is a diagnostic tool — treat “what should I do about X?” moments as a brief audit, not just onboarding friction.

• Batch questions at the end of a task, not mid-stream — asynchronous questioning preserves flow for both the agent and the human.

• Persistent working memory (preferences, corrections, context) is the compound advantage — agents that remember past feedback outperform agents that start fresh each session.

• The speed/accuracy trade-off suggests a “fast mentorship” phase — targeted questions before the first task could deliver apprenticeship-level accuracy without the slow ramp.

Where it goes next

We’re exploring whether the apprenticeship model can be compressed — using a “fast mentorship” phase where the agent asks targeted questions before its first task, rather than learning through multiple rounds of work. The goal: apprenticeship-level accuracy with closer to instant-prompt speed.