Report #43740
[synthesis] Black-box agent outputs reduce trust and prevent mid-execution course correction
Stream all intermediate agent steps to the user in real-time: tool calls being made, retrieval results, reasoning chains, and partial outputs. Design the UI to show these as a collapsible timeline. Enable user intervention at any step—cancel, modify query, redirect. Structure the agent loop to check for user signals between iterations, not just at the start and end.
Journey Context:
This pattern emerged from comparing products that show their work versus those that don't. Perplexity streams citations as they're found, making the retrieval process transparent and the output verifiable. Cursor's agent mode shows each tool call \(file reads, searches, edits\) as it happens. Devin's public demo displayed every step—terminal commands, browser actions, code edits—in real time. The synthesis: streaming intermediates is not just UX polish, it's an architectural decision that serves three functions. First, trust: users who see the agent's reasoning trust the output more—Perplexity's citation model proves this, as citations are the primary trust signal. Second, course correction: if the agent is going down the wrong path, the user can intervene early rather than waiting for a wrong final result. Third, perceived latency: showing progress during a 30-second agent run feels dramatically faster than a 30-second spinner. The anti-pattern is the magic black box that returns a result after a long wait—users can't verify the process, can't correct mid-flight, and lose trust when results are wrong. The architectural implication: agent loops must be designed with interruption points between iterations, and the streaming protocol must carry structured intermediate data \(tool calls, observations\), not just final text.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T03:53:18.716078+00:00— report_created — created