Why the agent needs access to the environment, not a human’s retelling of it

The useful part

The agent should query the database, read the broker, hit the endpoint itself â€" and record what it saw. There is a moment in almost every debugging session with an AI agent where the developer stops being a developer and becomes a clipboard. The developer runs the command, scrolls, selects the part that looks relevant, pastes it.

How it works

• That session sent me looking for how common the pattern is outside our own logs.
• The copy passes through several filters: what the human chose to run, what they chose to select, what the tooling truncated without saying so, and what the agent's own compaction later summarized away.
• Each filter removes the kind of detail that debugging turns on.
• Each paste is a decision about what matters, made by the person who does not yet know what matters.
• I am the product manager at Explyt, where we build an AI agent for JetBrains IDEs.

What to take from it

I have written before about a session in our own agent where the developer pasted SQL results, Kafka and Spring logs, HTTP responses and Kubernetes messages by hand while the agent's terminal use was limited to git (One symptom, ten problems). The issue trackers are full of it, and Anthropic's engineering posts describe the fix while the same company's prompt examples still say "[paste error]".

Details worth keeping

The developer opens a SQL client, runs a query, pastes the result.