AgentTalkie

Agents are multiplying faster than anyone can supervise them. AgentTalkie is one conversation across all of them.

It is a voice workspace for directing agents and tools, approving actions, and reviewing the tool call, result, and receipt in one place. The operator can follow work across systems without becoming the copy-and-paste layer between separate chats, dashboards, and terminals.

The recorded workflow

The public demo runs four integrations in one conversation:

1. Ambiguous AI drafts a Personal EA job description, waits for approval, creates a real document, and reads it back to prove the write completed.
2. Exa retrieves 10 public results in 1,006 milliseconds for $0.007, with the query, endpoint, and sources visible.
3. Codex through Ori builds a playable tic-tac-toe game in a fresh workspace while its native CLI output streams into AgentTalkie.
4. Claude Code through Ori edits the exact HTML artifact Codex produced and returns a new playable version.

This is the important handoff: Claude does not rebuild from a paraphrased prompt. It receives the exact parent artifact. Every successful version carries a content hash and parent reference, and the previous working preview remains available while an edit runs or if it fails.

System design

AgentTalkie separates a conversational request from the durable work behind it. Long-running coding jobs have explicit pending, running, failed, completed, and unknown outcomes. Tool activity and native output stay visible beside the conversation. Proposed writes are shown separately from completed actions.

• GPT-Live and WebRTC: a separate Live API surface using /v1/live/sessions, gpt-live-1, and client delegation so voice can hand work to external agents.
• CopilotKit and AG-UI: typed events for actions, progress, results, and receipts instead of parsed terminal logs.
• Ambiguous AI: one MCP across 17 agent-first apps, with live-schema validation before writes and readback afterward.
• Exa Code Context: source-backed public research inside the same thread.
• Ori, Codex, and Claude Code: local child processes using installed agent authentication, native CLI output, and exact-artifact handoffs.
• Next.js, React, and Postgres: the operator workspace, durable conversations, jobs, receipts, and artifact lineage.
• Docker: bounded runner filesystems and resources for coding jobs.

What I learned

• Voice is the interface, not the state machine. The transcript can steer work, but durable job state has to own what is actually pending, complete, or failed.
• Latency is part of the interface. An instant or fast search can stay inside a live voice turn. A 40-second deep-reasoning search feels broken unless the job leaves the audio loop and reports progress.
• Agent handoffs need identity, not summaries. A content hash and parent artifact ID are what let a second coding agent edit the right result.
• Local execution changes the trust boundary. Ori can inherit local agent authentication, MCP servers, and skills. That is powerful, but isolation has to be chosen and shown explicitly.
• Discovery is not permission. Seeing a tool in a catalog does not prove the current account can execute it.
• Writes need approval and readback. An uncertain mutation is not automatically retried, because duplicate side effects are worse than a visible unknown state.
• Failure should preserve the last good artifact. An edit can fail without taking the playable result away from the operator.

Current boundary

The recording is public. The live workspace requires an access code and configured providers. Codex and Claude Code are the verified coding harnesses; additional fleet adapters are not implied. Generated apps are self-contained HTML artifacts, not automatic repository deployments.

Read the build notes and technical lessons.