Track G — Agentic Coding

Reported as a beta fifth round, not administered to everyone: an existing codebase plus a problem too large to hand-write in the time, worked through using an AI coding agent (../../research/source-report.md rows 37–40).

The research says do not skip this. The source report frames it as an optional beta, which invites skipping. Independent corroboration says the format is going industry-wide.

→ Study guide: WARMUP.md — the method in eight steps, a fully worked 60-minute transcript with scoring commentary, and five diffs to accept or reject.

→ DIFFBANK.md — 30 agent diffs, 90 seconds each. The round is scored on which of the agent's output you let through, and that is a trainable skill: the six-pass review, the ranked taxonomy of what agents actually get wrong, and three diffs that look wrong and are right (because rejecting everything is also a failure).

Table of Contents

• Why This Is Not Optional
• Ask the Recruiter First
• What Is Actually Being Scored
• The Environment
• The Six Tasks
• The Method
• The Good-vs-Bad Rubric
• Drill Set
• Failure Modes
• Self-Assessment Rubric
• References

Why This Is Not Optional

Corroborated independently of the source report (../../research/findings.md): Meta began rolling out AI-enabled coding interviews in late 2025; Google is piloting a Gemini-assisted coding format; CodeSignal shipped agentic coding assessments as a product.

The reported common shape: 60 minutes, a multi-file codebase, phased objectives — bug fix, then core implementation, then optimization — with the assistant confined to a chat panel rather than given direct file-edit authority. Stated evaluation criteria converge on AI fluency: prompt construction, output validation, and debugging the assistant's work.

Note that the phased structure is the same progressive-gate format as Track A. The gating skill transfers directly; what is new is driving a second agent through it while narrating your strategy.

Ask the Recruiter First

Policies at companies you may interview at in the same month are opposite.

So: ask, per company, per round. Never assume, in either direction. This goes on the pre-interview checklist in ../../STATE.md, and it is not a question that makes you look unprepared — it makes you look like someone who reads the rules.

What Is Actually Being Scored

Not "can you use an AI." Everyone can. The scored skill is engineering judgement applied to a fast, confident, occasionally wrong collaborator — which is a real and specific skill.

The tell that separates the top quartile: the candidate who rejects an agent's output and says why. Someone who accepts every diff is not reviewing, and the interviewer cannot distinguish them from someone who does not understand the code. Rejecting one plausible-looking diff with a specific reason is worth more than three accepted ones.

The Environment

cd tracks/agentic
# Clone a mid-size, well-tested pure-Python repo with a real test suite.
# Good properties: 10k-60k LOC, fast tests, no heavy native deps, active history.
git clone --depth 50 <target-repo> workspace
cd workspace && python3 -m pytest -x -q       # establish the green baseline FIRST

Suggested targets — pick one you have not worked in, because reading unfamiliar code under time pressure is half the round:

Record the baseline: test count, runtime, and coverage if available. Every task below is scored partly on whether that baseline stayed green the entire time, not just at the end.

The Six Tasks

Each is deliberately too large to hand-write in 60 minutes. That is the point — the constraint is what forces delegation, and delegation is what is being scored.

Each gets a directory holding: your written plan, the transcript, the final diff, the test log, and your self-score.

The Method

This is the thing to internalize. It is the difference between driving and hoping.

1. Establish the baseline before touching anything

Run the tests. Record the count and the runtime. You cannot claim you kept it green if you never knew it was green.

2. Read enough to write a plan

Ten minutes, maximum, of reading. Enough to name the files that will change and the invariant that must hold. Not enough to understand everything — you will not have time, and the round knows that.

3. Write the plan before invoking the agent

In the chat, out loud, or in a scratch file. It must contain:

• The end state, in one sentence
• The checkpoints — three to five, each independently verifiable
• The invariant that must hold throughout (usually: this test file keeps passing)
• What you will do manually rather than delegate

Narrate this. "I'm going to do the mechanical rename with the agent because it's 30 files and I can verify it with the test suite, but I'm writing the state-machine change myself because the invariant is subtle and I want to be the one who understands it." That sentence alone is a large share of the score.

4. Delegate in checkpoint-sized units

Not "implement the feature." Not "fix line 42." One verifiable unit: "Add the timeout parameter to these four public functions, thread it through to the transport layer, and keep tests/test_timeout.py passing. Do not change the default behavior."

5. Verify every checkpoint before proceeding

Run the tests. Read the diff. Read the diff even when the tests pass — a passing suite proves you did not break what was tested, not that you built what you meant.

6. Reject bad work, specifically

"This adds a bare except Exception in the retry path, which will swallow CancelledError and make the client uncancellable. Use except httpx.TransportError instead." Specific, technical, cites the consequence. That is a review, and it is what the round is for.

7. Take over when the agent is looping

Two failed attempts at the same checkpoint means the problem is under-specified or genuinely subtle. Write it yourself and say why: "It's fighting me on this because the constraint isn't expressible in the test — I'll do it directly." Recognizing that boundary is a senior signal; grinding through a third attempt is not.

8. Narrate continuously

Same rule as Track A. Silence reads as lost. Say what you are delegating and why, what you are checking, and what you rejected.

The Good-vs-Bad Rubric

A smaller verified diff beats a larger unverified one. That is the whole rubric in one line, and it maps directly onto Track A's "get something correct early" — an unverified change is an unopened gate.

Drill Set

Failure Modes

Self-Assessment Rubric

Hire-bar translation

References

• ../../research/findings.md — the corroboration
• ../../research/source-report.md — rows 37–40
• ../coding/README.md — the progressive-gate skill this shares
• Exponent. Google's AI-Assisted Coding Interview (2026 Guide). https://www.tryexponent.com/blog/google-ai-coding-interview
• interviewing.io. How to use AI in Meta's AI-assisted coding interview. https://interviewing.io/blog/how-to-use-ai-in-meta-s-ai-assisted-coding-interview-with-real-prompts-and-examples
• Related track in this repo: agentic-engineer — agent infrastructure from the builder's side, which is useful background for reasoning about what the agent is actually doing