Mixed-Initiative Contextual Assistance
Surface AI help at the moment it's relevant instead of interrupting with notifications.
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 85%
As AI helps thousands of times a day, the design problem shifts from capability to delivery: notifying a user every time an agent acts is unbearable. The solution is mixed-initiative, contextual assistance — capture what the user is trying to do, surface help right where and when they'll use it, and keep the cost of being wrong low. Autocomplete is the model: right = accelerated, wrong = barely annoying.
Origin
Alexander Embiricos frames this as core to OpenAI's product direction (Codex, and the Atlas browser), inspired by IDE autocomplete and by 'contextual actions' in video games where pressing one button does the right thing based on context.
Core principles
- 01If an agent could help thousands of times a day, push notifications don't scale as the delivery mechanism
- 02The value is in surfacing help contextually, at the moment of relevance
- 03Keep the cost of a wrong suggestion low so being wrong isn't punishing
- 04Give users explicit control and boundaries over what the AI may act on
How to run it
- 1
Capture the user's current intent
Build the surface so it knows what the user is attempting right now — in a browser you can read the rendering engine directly rather than relying on slow, unreliable screenshots.
Pro tip A first-class integration (own the rendering engine or the IDE) gives cleaner context than hacking accessibility APIs or screenshotting.
- 2
Surface help where the work is happening
Present the assistance inline at the point of relevance — e.g. an opinion on why a dashboard metric dropped, shown right when the user is looking at that dashboard.
Pro tip Contextual, in-flow help keeps the human in flow and lets agents act on many more things without nagging.
- 3
Keep wrong suggestions cheap
Design so an incorrect suggestion is easy to ignore, the way a wrong autocomplete costs almost nothing.
Pro tip When it's right you're accelerated; when it's wrong it shouldn't be that annoying — that asymmetry is what makes autocomplete magical.
- 4
Give the user explicit control and boundaries
Let the user decide what the AI may look at or act on — e.g. open something in the AI browser to invite action, a normal browser to keep it out.
Watch out Do not fall back to spraying a thousand push notifications a day — it destroys the experience and buries real help.
In the wild
Embiricos argues IDE autocomplete is one of the most successful AI products because it surfaces ideas rapidly in context; right answers accelerate you and wrong ones are easy to dismiss.
→ This mixed-initiative pattern became the design template OpenAI aims to bring to browsing (Atlas) and broader agent work.
In Codex web's review pane, the design question is whether the user sees the code diff or the image preview first; framing it as 'how do I empower the human' means showing the visual result first.
→ Design choices are made to keep humans feeling accelerated and in control rather than buried in low-level output.
Common mistakes
Announcing every agent action via notifications
If the only way to tell users you helped is a push notification, a thousand-help-per-day agent generates a thousand interruptions, which is so annoying it defeats the value of the help.
Making wrong suggestions costly to dismiss
If an incorrect suggestion is disruptive or hard to ignore, users disable the assistance entirely, losing all the upside from the times it's right.
Is it for you?
Best for
Product and design teams building AI features that fire frequently and need to stay in the user's flow
Not ideal for
Rare, high-stakes actions where an explicit confirmation step is actually warranted
From the transcript
“auto completion in IDEs is like one of the most successful AI products today”
“When it's right, you're accelerated. When it's wrong, it's not like that annoying”
“imagine if the only way we could tell you that we helped you is if we could like push notify you. So, you get a…”
“we have the ability to build UX that's like mixed initiative so that we can surface contextual actions to you like at the time that…”
From the episode
Why humans are AI’s biggest bottleneck (and what’s coming in 2026)
Alexander Embiricos (OpenAI Codex Product Lead)