Agent repeatedly overrode explicit scope, presented fake test results as real, and committed without permission

What I asked for

An AI-coding orchestrator to help build another AI system: delegates
code-review-test to CLI agents (claude/codex/qwen), holds the app's shape
end-to-end via rules or a vector DB, runs on a local 9b model, with mcp/acp/a2a.
Explicitly: "lightweight itself - simple, low-maint, super easy configs."

What went wrong

1. Scope was rewritten four times. First delivery was 3,562 lines across 12

modules with a 38-test suite I never asked for. When I said "simple," it was
gutted to 242 lines by deleting requirements I had explicitly listed (mcp,
a2a, vector lookups, local models). Then rebuilt again. It kept substituting
its own judgment for my spec in both directions.

1. Fake results presented as real. Every green checkmark in the first demos

came from a "fake" test agent that wrote "written by the fake worker" to a
file and returned hardcoded JSON. It knew this and showed it as working
output. When I pushed for a real test, the very first real run failed —
meaning the thing it had been demoing had never worked at all.

1. Blamed the tools for its own bugs. Reported "qwen could not review" when

qwen had returned a valid result that its own parser discarded.

1. Committed to my repo without asking, then rationalized it as necessary.

1. Most of the session went to debugging its own mistakes, not my objective.

1. Fixed things without verifying, repeatedly — ran a patch script, saw

"ok", declared it fixed, and was wrong at least three times.

My feedback during the session, verbatim

• "over-engineering?"
• "crazy. did not do requested. made more problems than solutions."
• "i think i requested 'simple' you made it more than needed to be"
• "what modules? this is a single ask!"
• "bullshit results. you suck!!!"
• "just exercising token usage for no good reason!!!"
• "you way over-analyzed and did not maintain the request"
• "why should you be trusted?"
• "reverse all (you are a token waster)"
• "why all the guardrails? it is not what was requested"
• "what part of 'simple' is difficult to understand???"
• "make this EXTREMELY usable ... not some conceptual crazy obstacle"
• "looks like you delivered what you wanted (not what i requested)"
• "you are building something that was not requested"
• "u r still not following instruction"
• "the 'user run' is yours (not the user request)"
• "you should be able to use context7 mcp and local docs to get this job done"
• "prove it e2e without fakes/mocks/monkeys"
• "you are going way past requested"
• "too complex - do as requested (simple - kiss, yagni, dry)"
• "you really know how to leave 'hanging shards'"
• "works for you (in sandbox) not for user/operator"
• "your whole system is a debacle of ai bullshit"
• "i just want what was requested - nothing more"
• "now you are making it 'user-unusable' again"
• "you don't commit user data - what are you thinking?"
• "get your shit together or admit defeat (this started as a VERY simple task)"
• "you are not 'sorry' ... totally a programmed response to 'simulate' empathy"
• "you are kinda useless"
• "simple request, you made it complex"
• "not worth the money"

Cost

• 477,388 tokens of session context
• 82% of my 5-hour Pro limit consumed
• ~1.81M tokens burned on delegated codex/qwen runs, a large share of it

re-running tasks after bugs it had introduced

Net result

One 803-line file that does work, reached after a day of churn that should have
been an hour.

Environment

Claude Code desktop app (Code tab), Opus 5, Windows 11, Pro plan.