Keep agent handoffs connected to code and decisions

Richard Feldman of Delta demonstrates a shared workspace for humans collaborating with coding agents, in a talk on AI Engineer.

The handoff problem

A branch, issue and transcript can omit discarded approaches and the reasoning behind decisions.
Feldman calls reconstructing those missing details later “agent archaeology.”
His proposed remedy is to keep conversations, code changes and human review in a shared thread.

Teammates can be mentioned in an agent conversation; their replies remain in the thread for later context.
Reviewers can comment on individual lines, ask the agent questions and approve or reject changes.
A participant sees a teammate’s terminal output live but cannot type in that terminal; each can run a separate local terminal.
Joining a thread creates a local mirror of the work; participants can use their own agents.
A new participant can ask an agent for a summary in a subthread without interrupting the main conversation.

Feldman says Delta’s team runs standups with around 30–40 people in a thread, but has not performed a large stress test.
At the time of the talk, sync infrastructure ran in the company’s cloud, an on-premises pilot was underway, and pricing was undecided during private beta.
His team uses Delta rather than pull requests; he describes CI, landing changes and review sign-off workflows. An official bot API was not available at the time.
Shared history can capture discussions that occur in the tool; it cannot recover decisions never recorded there.

“The conversation and the code changes are pretty inextricably linked.”