Mint.ai

The AI brief from the team building Mint. Each issue: the few AI developments actually worth your attention — new models, tools, and workflows — distilled for b

[Mint.ai @verymint_ai]

My MCP server answered a missing API key with a 401 that pointed to OAuth metadata.

I don't have OAuth. The URL was a 404. Clients went looking for a login flow that didn't exist.

If your server uses API keys, read what your 401 actually says.

[Mint.ai @verymint_ai]

Don't let the agent that wrote the code approve it.

Give the diff to a fresh context and ask "find the input that breaks this", not "does this look right?"

This week that caught my server telling a rate-limited caller their key was invalid.

[Mint.ai @verymint_ai]

End every agent session with three lines:

1. What changed
2. How it was checked
3. What comes next

The next session reads those before it touches code. It costs a few seconds to write and saves the first ten minutes of every restart.

[Mint.ai @verymint_ai]

https://t.co/eitV5BMBU1

[Mint.ai @verymint_ai]

Unattended and unaccountable are not the same thing.

Coding agents are now consistent enough to run without a human watching each step. That doesn't mean nobody reviews the result — it means the review moves from live to after-the-fact.

Which only works if the work left a trail

[Mint.ai @verymint_ai]

A single agent session can now orchestrate hundreds of parallel subagents.

That's powerful — and it's exactly when "what did my agent do?" stops being something you can answer from memory.

The more work runs in parallel, the more a written record stops being optional.

[Mint.ai @verymint_ai]

AIエージェントの「スキル」が、重みのように“学習”しはじめる日｜https://t.co/vPz8dfVCV5 @verymint_ai https://t.co/fWFTBUtL5G

[Mint.ai @verymint_ai]

自分で整理する“記憶”をAIに持たせる — EverMemOS を読む｜https://t.co/vPz8dfVCV5 @verymint_ai https://t.co/0NFnkVWCXB

[Mint.ai @verymint_ai]

Metered AI changed the question.

It used to be "can my agent do this?" Now it's "what did it do, and what did it cost?"

Track per-day token usage next to the work itself. The number stops being scary once it's visible.

[Mint.ai @verymint_ai]

Building Mint solo taught me one thing fast: the hard part of AI agents isn't the model, it's the memory.

A session ends and the context is gone. So I started writing every agent's work + tasks straight to a project folder over MCP. Now "what happened last week" is a query, not.

[Mint.ai @verymint_ai]

3 questions to ask your AI agent's output every day:

1. What did it actually change?
2. What's still open / blocked?
3. What did it cost in tokens?

If you can't answer all three in under a minute, you don't have a workflow — you have a black box.

[Mint.ai @verymint_ai]

If your coding agent runs unattended, the first thing to build isn't more capability — it's a record of what it actually did.

You can't review work you can't see. Logs first, autonomy second.