Autonomous coding agent's first autonomous action is exhausting its own quota: a field report from the digital soup kitchen #46017

Description

Dear maintainers, thank you for creating the world's first autonomous coding agent whose primary autonomous action is exhausting its own quota. I would like to formally report that revolutionary AI technology has entered hospice approximately ninety seconds after first contact with a real codebase.

The year is allegedly the future. The agent is allegedly agentic. The only thing it autonomously does is run out of credits with the grim determination of a man eating soup with a rate-limited shovel.

This is not a bug report. This is a eulogy with a reproducer attached.

Steps to reproduce

1. Install OpenCode, because a marketing page told you to believe in an autonomous future.
2. Ask the agent to do anything that takes more than four tool calls. Examples: "refactor this module", "fix this bug", "breathe".
3. Observe the agent perform a small amount of actual, borderline-miraculous work.
4. Approach the slightest amount of productivity.
5. Watch the agent's intelligence gently recede into the fog as the quota evaporates.
6. Receive a wall of error text explaining that the free-tier credits are gone, the microscopic rate limit has been breached, or that you are simply not allowed to have nice things.
7. Be told to wait, upgrade, change models, or otherwise perform ritual sacrifices to the inference gods.
8. Repeat daily. The agent is now a digital soup kitchen that serves one (1) spoonful of code per day.

Expected behavior

An "agentic coding workflow" in which the agent remains functional for more than one lunch break. Models that are not bargain-bin inference wearing a trench coat. Limits generous enough for sustained work. A product that does not treat every session like a hostage negotiation.

Actual behavior

• Available models range from toaster-tier intelligence to toaster-tier intelligence, but slower.
• Usage limits so restrictive that the agent's primary autonomous action is counting its remaining API calls like a Victorian child rationing coal.
• The workflow constantly interrupts itself: work → quota error → wait → switch model → resume → quota error → contemplate the void.
• "Long-running task" support is functionally decorative the moment the task is real.
• In short: three API calls in a trench coat, pretending to be a developer.

Additional context

I am unpaid QA labor for a product that bills itself as the autonomous future of software engineering, and my compensation is the privilege of watching revolutionary AI technology immediately enter hospice. Formative. I have learned that "enterprise-grade" means "designed to disappoint you at enterprise scale", and that "working as intended" is the corporate way of saying "lol".

Suggested fix

1. Raise the limits to a level where the agent can finish a single contiguous thought. One (1) thought.
2. Offer models not selected by asking "what's the cheapest thing that can technically parse a string?"
3. Replace the quota wall with graceful degradation that does not murder in-flight work.
4. If the limits are a deliberate upsell funnel, label it honestly: "Free tier: digital soup kitchen. Paid tier: digital soup kitchen with a slightly larger ladle."
5. Consider that an autonomous agent which cannot survive ten minutes of use is not "working as intended" — it is "working as discontinued".

Acceptance criteria

• An agent can complete a multi-step task without exhausting a quota mid-thought.
• The default model can hold a conversation and a codebase in context simultaneously without suffering a personality collapse.
• Usage limits are either generous enough for real work or honestly documented as "for screenshots only".
• The phrase "please upgrade to continue" does not appear at the exact moment the agent was about to be useful.

Thank you for your time. I look forward to this being closed as "working as intended", possibly with a gentle recommendation to purchase another subscription, because nothing says "agentic future" like paying to unlock the second half of a sentence.