o1 and Gemini 2.0 pushed agents toward the default

o1 and Gemini 2.0 made agent-shaped product work harder to avoid.

o1 made reasoning time visible for difficult math, coding, and multi-step tasks. Gemini 2.0 pushed multimodality, tools, and agent prototypes into the product story. Together, they made the chat tab look like one surface inside a larger work loop.

Longer reasoning creates a routing problem. A spelling fix should not receive the same budget as a migration plan. The product needs controls for effort, latency, interruption, and review.

Tool use makes the boundary concrete:

• which tools exist?
• what can each tool read?
• which arguments are allowed?
• what needs approval?
• what gets logged?
• what may the model retry?

Once those questions exist, the interface includes permissions, run history, approvals, results, and recovery.

Multimodal state widens the workspace but also the evidence problem. A screenshot, browser state, PDF, or video can speed diagnosis while introducing OCR, cropping, transcription, and interpretation errors. Record what the model observed directly, what was extracted, what the user claimed, and what the user can inspect before approval.

Agent-shaped does not mean unsupervised. The loop can remain cautious:

understand task
  -> gather context
  -> propose step
  -> use tool or request approval
  -> inspect result
  -> continue, stop, or escalate

The product needs receipts: requested task, context, tools, arguments, changes, evidence, approvals, failures, and rollback.

Evaluation moves from answer quality to run quality. Did the agent choose the right tools, respect scope, stop at the success condition, recover from failure, and leave a clean workspace?

o1 and Gemini 2.0 did not make reliable autonomy inevitable. They made the infrastructure around multi-step work impossible to ignore.