Talking to Agents: The Missing Context
We need different communication patterns with AI to unlock what these systems can do.
2026-01-15
Loading content...
We need different communication patterns with AI to unlock what these systems can do.
One-shot prompts treat the agent like a vending machine. The useful sessions look more like talking to a colleague: you bring screenshots, you answer follow-ups, and the thread keeps state so you're not re-explaining the repo every message.
Last month I tried to fix a checkout bug by pasting "payment is failing, fix it" into Cursor with the whole repo indexed. The agent wandered through CSS, CI config, and a 2019 inventory sync before touching Stripe. Same mistake I'd make briefing a human intern badly, except I know how to brief humans.
With my actual PM I send the error, the file I suspect, what I already tried. With the agent I kept acting like the prompt box was a shell command.
I stopped optimizing single prompts and started optimizing sessions.
On a gnarly auth bug I did what I'd do with a teammate: pasted the Sentry stack trace, @'d middleware.ts and the session cookie helper, said "I already checked env vars." The agent asked whether we were on App Router or Pages (fair question). I answered. Third message it had a patch. Not magic. Just the same back-and-forth I'd have on Slack.
That maps to how I already work elsewhere:
The gap isn't model IQ. It's that most UIs still feel like a one-shot command line with a chat skin.
I'm not waiting for "agentic AI" as a buzzword. I want interfaces that assume ongoing work:
Thread memory that survives a day. If I debugged checkout yesterday, today's message shouldn't need the Stripe callback story again.
Attachments that carry meaning. A Figma frame or a screen recording should land as "this spacing is wrong on mobile," not a blob the model pretends to read.
The agent pushing back. "Which environment?" and "Did this break after the deploy?" save me from vague one-liners.
World outside the chat. Link the Linear ticket, the prod error rate, the design doc so I'm not the integration layer every time.
Cursor, Claude Code, and Copilot are inching there. None of them fully replace "brief a sharp intern with the right three files." Until they do, I'll keep over-investing in the first three messages of a thread. That's where the quality actually lives.