← All pages

Pushing Back on AI Coding Agents

Twenty practical ways to challenge an AI coding agent, grouped by where the gaps show up. Guidance only: each point is something to do, not something to know.

Verify claims before accepting them

  1. 1

    Ask for evidence, not assurance.

    “Tests pass” and “done” are claims. Ask for the command, the output and the exit code, for every claim.

  2. 2

    Ask “what did you not check?”

    Agents report what they did, not what they skipped. If an agent finds no PR comments through the API, ask which sources it queried.

  3. 3

    Demand a negative-result report.

    “Found zero X” must come with the endpoints or paths searched. Make “zero results, here is where I looked” mandatory.

  4. 4

    Test the fix, not the agent’s story about it.

    Re-run the failing case yourself, or ask for a failing test written first. A diagnosed fix that was never committed or verified is not a fix.

  5. 5

    Check the diff, not the summary.

    Read git diff --stat before accepting. A summary says what the agent intended; the diff shows what it did.

Control scope

  1. 6

    Ban blanket staging.

    git add -A can sweep agent worktrees into a PR. Require explicit paths and enforce it with a PreToolUse hook, not a polite request.

  2. 7

    State the write boundary up front.

    “Only touch files under X” prevents stray planning files. Agents expand scope unless fenced.

  3. 8

    Ask for the plan before any edit on more than about 3 files.

    Review it for what is missing (migrations, callers, docs, tests), not just for what is present.

  4. 9

    Challenge the size.

    If the diff exceeds your PR cap, ask for a split proposal before work continues. Oversized PRs happen because nobody challenged size early.

  5. 10

    Stop scope creep with a “why” test.

    For each changed file ask “which requirement needs this?” Revert what cannot answer.

Question the decisions

  1. 11

    Ask for alternatives and when each wins.

    “What are two other approaches and when would you pick them?” Agents default to the first plausible one.

  2. 12

    Ask what breaks.

    “How does this fail in production?” surfaces thresholds, race conditions and rollback gaps the agent glossed over.

  3. 13

    Challenge unexplained numbers.

    Rollout steps, error thresholds and timeouts with no rationale are guesses. Insist on a source, a derivation, or the label “starting point”.

  4. 14

    Push on tool defaults.

    Agents bake in popular tools as if they were the principle. Ask “what is the requirement, and is this tool one example?”

  5. 15

    Check for hallucinated specifics.

    Verify that versions, flags, API fields and file paths exist, with --help, a doc lookup or ls. Fluent wrongness is the agent’s signature error.

Manage the process

  1. 16

    Cut loops at two repeats.

    If the agent says the same thing twice, it is stuck. Interrupt and change the instruction, the hook or the state; do not wait it out.

  2. 17

    Fix the system, not the output.

    When a correction repeats (BSD sed, git add -A, stray plan files), put it in CLAUDE.md or a hook. Correcting the same mistake three times in chat is a process failure.

  3. 18

    Separate permission from capability.

    When a classifier blocks a destructive action, ask for the exact command block to run yourself. Do not let the agent route around the block.

  4. 19

    Give intent with terse commands.

    A bare “plan” produces nothing. Add scope, inputs and done criteria: what, where, and how you will know it is finished.

  5. 20

    Make the agent state assumptions and confidence.

    Ask “what did you assume, and what is least certain?” before big actions. Compacted sessions lose context, so verify state (git status, branch, open PRs) instead of trusting the summary.

What people typically miss

Interview framing

Describe pushback as a loop: bound the task, demand evidence, check the diff, probe alternatives and failure modes, then encode the lesson in a hook or standard. Anchor each step in a real example: a missing-comments case, a blanket git add -A incident, or a CI check that replaced a judgment call with an automated rule.