Richard Feldman: keep the agent conversation with the code

A shared thread can preserve decisions, reviews and handoff context that otherwise gets lost between teammates.

RFRichard Feldman登壇者 · Delta
チャンネル
AI Engineer
公開日
2026年11月9日
動画
19:24
読了時間
3 分
原文
英語
YouTube で開く
スクロールすると動画は右下に小さくなって再生を続けます。字幕は閲覧言語に合わせて切り替わります。
閲覧言語2 言語で読めます
この記事はまだ日本語に翻訳されていないため、英語の原文を表示しています。
要約

30 秒でわかる

Handing a teammate a branch and an agent transcript may leave out failed approaches and the reasoning behind decisions. Delta’s Richard Feldman calls reconstructing that context “agent archaeology” and demonstrates his team’s shared workspace for conversations, code review and real-time collaboration. Colleagues can join an agent thread, inspect changes and ask the agent to summarize earlier work. The product was in private beta at the talk; Feldman says his team has tested 30–40 people in one thread but has not established a maximum.

  1. 1A branch and transcript leave important handoff context behindFeldman says failed attempts and decisions may never appear in commits or the copied conversation.
  2. 2A teammate’s answer can become context for the agentHe describes a colleague asking about an old feature inside the shared work thread.
  3. 3Review comments live beside the code and agent discussionThe demo attaches line comments to changes and supports teammate review and edits.
  4. 4Teammates can watch terminal output without controlling the remote terminalFeldman demonstrates live output and a separate local checkout.
  5. 5The team tested standups with 30–40 people in one threadFeldman says no large stress test had established a definitive limit.
  6. 6A newcomer can ask the agent for a summary in a subthreadThe main thread continues while the newcomer asks follow-up questions.

ポイント

  1. Preserve decisions and failed approaches alongside the code when handing work over.
  2. Let teammates ask questions in context that their agents can also access.
  3. Attach review comments to the relevant code and retain the discussion.
  4. Use separate subthreads to catch up without interrupting the main conversation.
  5. Treat product scale, deployment and pricing as details to verify for your own rollout.
全文

チャプターごとに整えた全文を読む

文字起こしのみ
01

Why handoffs lose the reasons behind code

Feldman describes missing failed attempts and decision context as agent archaeology.

Richard Feldman asks listeners to imagine a coworker leaving for vacation after sharing a branch and a transcript of an agent conversation. The recipient has the issue and some code snapshots, but may not know which approaches failed or why the coworker made particular choices.

Commits do not necessarily preserve abandoned attempts, and a pasted transcript may omit other discussions or reasoning. Feldman compares the handoff to a puzzle with missing pieces: the recipient and their agent must reconstruct context before continuing.

Months later, someone extending the feature may have to infer why a decision was made from commit messages and fragments. Feldman calls this agent archaeology: recovering context that existed during the work but was not retained alongside it.

02

Put people, code and agents in one thread

A coworker interaction illustrates how the shared conversation can retain a decision.

He recounts a colleague mentioning him in Delta, the collaboration tool his team builds, to ask about an old undo feature. Feldman explained that his work was stale and the colleague could take it over. Their exchange remained in the agent thread, so the agent could use the answer as context.

Feldman argues that code changes, agent conversations and human discussions belong together. In Delta a teammate can see a diff, join the discussion and work with an agent. He demonstrates the product, which his team also developed.

03

Review and collaborate in Delta

The demo covers mentions, code review, shared terminal output and local worktrees.

The demo shows a macOS application, with Linux, Windows and mobile versions also mentioned. Feldman highlights text from an agent conversation and asks a follow-up question. He says the product aims to work with different models and agent frameworks.

He mentions a coworker, Bennett, in the thread to ask whether an error should be a toast or an inline message. Bennett joins and chooses inline. The decision becomes part of the shared conversation for the current agent and future collaborators with access.

Feldman shows a line-by-line code review. Comments can be attached to individual lines and sent to an agent; teammates can also review, approve or reject changes and directly edit code. Threads and nested subthreads provide separate contexts.

Bennett invites him into another thread during the presentation. Feldman can see Bennett's terminal output synchronized in real time but cannot type into that terminal, which would grant remote execution on Bennett's machine.

Feldman opens his own terminal, runs the project locally and says joining the thread created a local mirror of the repository. He calls these Delta worktrees rather than Git worktrees. Each participant can work locally and choose an agent while sharing context.

He returns to the handoff: a shared history can preserve agent conversations, code changes and human decisions together, making the evolution of work easier to inspect than a branch and scattered snapshots alone.

04

Questions on scale, deployment and workflow

Feldman discusses limited scale testing, hosting, pricing, CI and catching up on a thread.

Asked about scale, Feldman says the team has used a single thread for standups with roughly 30 to 40 people, but has not performed a large stress test and cannot give a definitive limit.

The demonstration uses cloud-hosted synchronization, while agents can run locally or on cloud runners. Feldman says an on-premises pilot is underway. At the time of the talk, Delta was in private beta, pricing had not been published and the beta was free.

On the development lifecycle, he says his team has stopped using pull requests internally: an agent can run CI and then land a change, and a formal approval workflow is available. He says a bot API was not yet officially available.

Asked how a newly invited person catches up, Feldman describes asking the agent for an overview in a subthread. The original conversation can continue while the newcomer asks follow-up questions separately.

同じチャンネルの記事

スライド 7 枚19:41EN
AI Engineer· 2026年11月9日A high-reliability automation starts by routing the hard cases awayMatthew Jewkes · Standard CyberneticsMatthew Jewkes describes separating critical errors, ambiguous decisions and unusual exceptions in a biopharma indexing workflow.エージェント工学Workflow automation評価(Evals)+1
スライド 7 枚22:01EN
AI Engineer· 2026年11月9日Zach Lloyd: make the software factory measurable and editableZach Lloyd · WarpWarp’s CEO proposes recording agent runs, scoring failures and testing changes to the workflow on real tasks.ソフトウェアファクトリーコーディングエージェント評価(Evals)+2
スライド 12 枚27:35EN
AI Engineer· 2026年11月9日Uber traded an always-green trunk for faster AI-scale mergesPreetam Dwivedi, Manjari Akella · UberIts merge queue now relaxes some conflicts and validates the combined state after landing, with a verified-good fallback.開発ワークフローとテストコーディングエージェントインフラとデプロイ+2
スライド 3 枚15:25EN
AI Engineer· 2026年11月9日Hassan and Omar El Mghari: When an AI financial forecast is right for the wrong reasonsHassan El Mghari, Omar El Mghari · Together AIA Costco membership-fee model passed basic checks and landed close to reported results, but an audit uncovered offsetting assumptions.金融 AI評価(Evals)エージェント工学
スライド 3 枚14:38EN
AI Engineer· 2026年11月9日Jemima Abu: Give coding agents accessibility rules from the startJemima Abu · CAISA product engineer shows how semantic HTML, reusable components and explicit agent skills can help teams build more usable interfaces.コーディングエージェントプロダクトと UXAgent skills
スライド 4 枚21:12EN
AI Engineer· 2026年11月9日Tom Moor: Give coding agents the product context that code cannot showTom Moor · LinearLinear’s workflow carries customer requests, past decisions, review history and human feedback from intake through implementation and review.コーディングエージェントプロダクトと UXエージェント工学