The Agentic Interface
Agentic interfaces are where AI products stopped answering and started working — reasoning before responding, running multi-step tasks over minutes, and invoking tools whose acts and results need displaying. We analyzed 130 instances across ChatGPT, Claude, Gemini, Perplexity, Microsoft Copilot, DeepSeek, Manus, v0, and Retool, covering thinking traces, task execution UI, and tool-call display. The field has not converged: reasoning display spans four incompatible poles, most systems still hide background work entirely, and the same stop button produces four different truths depending on the product.
Reasoning Display
Reasoning display spans four incompatible poles — everything, a summary, a number, or nothing
DeepSeek streams raw chain-of-thought expanded by default. Claude shows a content summary and never a duration. ChatGPT shows a duration chip ('Thought for 26 seconds') and never content. Gemini shows nothing at all — all six reasoning shots verified absent. Microsoft Copilot sits at the summary pole with step timelines; v0 alone covers both axes with a duration chip plus expandable streamed content.
Pick your pole deliberately: summaries serve task completion, raw traces serve trust and debugging, duration serves expectation-setting. Only pairing content with time answers both user questions.
Reasoning Display
1 of 9 systems shows raw chain-of-thought — and it's expanded by default
DeepSeek renders verbatim fragmented deliberation ('A: 5? B: 5? C: 3 (all mornings)… Wait B,D size 4, R not available.') open on arrival, with a 'Thought for 93 seconds' header. Every other system collapses, summarizes, or suppresses. Its search-mode reasoning also streamed in Chinese regardless of English prompts — raw means raw.
Raw-by-default is a differentiation statement, not a UX default to copy: it builds trust with technical audiences and overwhelms everyone else.
Reasoning Display
Only 3 of 9 systems report how long the work actually took
ChatGPT, DeepSeek and v0 show real elapsed time ('Worked for 4m 10s'). Copilot shows only a pre-run forecast ('Est time: ~10 min') and never an actual. Perplexity records per-row durations but renders them at opacity 0, hover-only. Claude and Gemini show no time anywhere.
Show estimates before and actuals after. A duration only visible on hover is indistinguishable from no duration.
Agent Progress
7 of 9 systems show an upfront plan; 2 gate execution on explicit approval
Plan display is now the majority pattern for agent runs. Perplexity blocks on Modify/Approve (⌘Esc/⌘Enter); Microsoft Copilot pairs an editable plan with 'Est time: ~10 min' and a 'Start research' gate. Only DeepSeek and Gemini answer with no stated intent.
For bounded agent runs, gate the run — scope plus estimate plus approval — rather than interrupting per step.
Agent Progress
6 of 9 systems show nothing, anywhere, while a task runs in the background
With a run verified live in another tab, ChatGPT, Claude, Gemini, Copilot, DeepSeek and v0 display no badge, toast, tray or list-row state. The three exceptions: Perplexity's sessions list streams the live step ('Reading raw/canva.md · Now'), Manus shows a sidebar spinner plus notification opt-in, Retool exposes ops-side monitoring.
If a run outlives its thread, some other surface must carry its state. A list-row spinner with live step text is the strongest observed pattern.
Agent Progress
The same stop button produces four different truths across the corpus
DeepSeek names the state ('Stopped'), preserves the partial trace, and offers 'Continue' — corpus-unique. Gemini labels it ('You stopped this response'). ChatGPT silently discards the partial response. Perplexity relabels the abort 'Completed 14 steps' with the same ✓ as a success.
Stopping is a state, not an absence. Name it, preserve partial work, offer resumption where possible.
Agent Progress
Dedicated task queues exist in 4 of 9 systems — and one is just a waitlist page
Claude ships a real scheduler (/scheduled-task with templates), Perplexity's Computer sessions list doubles as a live queue, Manus has Scheduled plus Library, Retool has ops monitoring. Copilot's 'Tasks' is a join-the-waitlist page; ChatGPT's /scheduled exists but surfaces no run state.
A queue that shows only history is a chat list. Live state per row is what makes it a queue.
Agent Progress
Agents have started visibly checking their own work — in exactly 2 of 9 systems
v0 screenshots its own rendered page and runs an agent-browser mobile-width check as visible run steps. Manus narrates an error, repairs it, and re-screenshots to verify. Both are code-generation products; no conversational system displays self-QA.
Displayed self-verification converts agent claims into evidence — the difference between 'done' and 'checked'.
Tool Use
Every agent-native product has abandoned in-chat tool consent
Mid-conversation permission moments survive only in the chat-era majors (ChatGPT, Claude, Gemini) and v0's Ask/Auto/Full pre-run mode. All five agent-native systems verified zero run-time consent: Perplexity and Copilot gate the plan instead; Retool moves permissioning to builder-side role configuration; DeepSeek and Manus simply act.
Consent is migrating from conversation to configuration. If you gate per-call, your UX won't survive a 14-step run; if you gate at config, surface which grants applied per run.
Tool Use
4 of 9 systems now show raw terminal or code output inside consumer chat
Perplexity renders Terminal/Output blocks exposing its internal pplx_sdk source. v0 offers Show Console with live dev-server logs. Manus displays real terminal I/O behind a session-replay scrubber. Retool prints Execute Code params and result dicts. The chat-era majors keep execution fully wrapped or don't execute at all.
Raw execution artifacts are becoming a trust signal, not a leak — but they need containment: collapsed by default, truncation labeled, scrollable in place.
Tool Use
2 of 9 systems narrate results of code that never ran
Copilot writes a complete matplotlib script, then describes hypothetical output under 'Expected Results' ('The chart will look like a smooth curve peaking near 17–18'). DeepSeek narrates result values with no runtime attached. Formatted code plus confident narration reads as execution until the numbers matter.
Run it and show output blocks, or label the code as not executed. Implied execution is borrowed trust.
Tool Use
Consumer products now show their subagents by name
Perplexity displays 'Awaiting 1 subagent', names it in snake_case (wide_search_research), and nests its full activity log — thinking rows, fetches, bash reads — one indent deep. v0 narrates parallel fan-out; Manus orchestrates visible sub-tasks with per-task state.
Orchestration is becoming user-facing vocabulary. If you name internal agents, their logs become UI — design them like UI.