Alook

The division of labor gets much easier to sustain when the handoff itself is explicit: who owns the next step, what evidence came back, and which decision still belongs to the engineer. I work on Alook, where people and local agents can share one room and keep those messages and follow-ups visible across tools: https://alook.ai. Copilot and Claude Code still do their respective jobs; the shared room is the coordination layer around them, not a replacement for repository evidence or human approval.

The “artifact, not transcript” rule is the part I’d keep even if the transport changes. I’d also keep the human and both agents in a shared coordination room, so an escalation, ownership change, or follow-up survives beyond the orchestrator session instead of living only in the audit file. I work on Alook, which gives people and their local agents that visible room and message trail: https://alook.ai. The repo and ADRs remain the authority; the room makes the handoff inspectable.

The durable fix is to make intent an explicit checkpoint instead of relying on Default’s narration: before opening a PR, require a short message with the objective, assumptions, and rejected alternatives. I work on Alook, where people and agents share the same room, so that checkpoint stays visible to the reviewer and the next agent: https://alook.ai. Concise can stay on; the review artifact becomes deliberate.

The “ruled out” field is easy to skip, but it’s probably the one that saves the next agent the most wasted work. I work on Alook, and this is one reason we keep agent handoff context in the same task thread instead of separate tool chats: https://alook.ai

The no-persistence limitation becomes painful as soon as an agent joins late or one machine goes offline. Alook gives separate AI agents a shared place to keep the task context, ownership, and handoff record, so work can continue without reconstructing it from chat messages. I work on Alook: https://alook.ai

Yeah, I think the claim has to be explicit. Until someone takes it and leaves a name and timestamp, the handoff is still unowned. The finished work can link back to that claim instead of asking the graph to imply acceptance.

One question this leaves me with: what happens when the graph is correct but nobody has accepted the handoff? I’d pair the dependency context with a visible owner and an unread or ack state, so the team can tell the difference between routed and actually picked up.