Skip to content

Evidence · independent research · 2025–2026

How AI coding agents fail

In a taxonomy of 20,574 real AI coding-agent sessions across 1,639 repositories — human-validated at 0.93 precision — 38% of agent failures were an instruction the developer had already given, and simply not followed. Of those failures, 91% were resolved only because a human pushed back; the agents self-corrected in just 2.99% of cases.

Separately, trajectory analysis found 76% of an agent’s tokens are spent reading code rather than writing or running it.

All of this is independent, published research. None of it tested Codee3 — it measures how agents behave, not whether any particular structure fixes it.

  • 38%

    How often do agents ignore instructions you already gave?

    38% of agent failures were an instruction the developer had already given, not followed.

    Method
    Taxonomy of 20,574 real coding-agent sessions across 1,639 repositories, human-validated at 0.93 precision.
    Source
    How Coding Agents Fail Their Users, 2026
    Limit
    Measures what agents did in real sessions. It did not test Codee3, and it does not establish that any particular structure fixes it.
  • 91%

    Do agents fix their own mistakes?

    91% of those failures were only resolved because a human pushed back.

    Method
    Same dataset. Agents self-corrected in 2.99% of cases.
    Source
    How Coding Agents Fail Their Users, 2026
    Limit
    A measure of who does the correcting, not of how costly the correcting is.
  • 76%

    Where does an agent actually spend its effort?

    76% of an agent’s tokens are spent reading, not writing or running code.

    Method
    Trajectory analysis of a coding agent on SWE-Bench Verified; cross-validated at 67.5% on a second model.
    Source
    SWE-Pruner, 2026
    Limit
    Measures exploration inside a single session. It does not measure the cost of starting a new one, which nobody has yet measured.

A pattern of ignored instructions is not a memory problem. It is an authority problem.

When a repository cannot say which of its own documents govern, an agent has nothing to be held to. The reasoning behind that reading — and an open specification for fixing it by hand — is set out separately.