The coding agent session stopped being a private conversation today.

Juggler is a new open-source GUI coding agent from Julian Storer — the creator of JUCE and the Tracktion DAW, thirty years of shipping C++ audio tools — and its two design decisions are the interesting part. First, sessions are "a tree, not a doom-scroll": conversations branch into sub-threads you navigate in a Finder-style Miller column view, instead of one linear transcript scrolling into the void. Second, the session is genuinely multi-client — a desktop app (macOS, Windows, Linux) and a headless CLI server ship together and stay synchronized, so several clients, browsers, or remote machines can attach to the same live session at once. The synchronization layer is Yjs — the same CRDT machinery under multiplayer document editors — which tells you what a session is here: a shared document, not a process's stdout. Model-agnostic (Claude Code, OpenAI, Gemini, Ollama, OpenRouter, Deepseek), extensible without rebuilding the core. AGPL-3.0 app, Apache-2.0 extensions, 367 stars, created four weeks ago, 244 points on Show HN.

Reach for it when the linear transcript is the thing losing you — when a session's third tangent deserves its own thread, or when you want the session running headless on the box where the code lives while you watch from a laptop. Delete the tmux-plus-prayer setup for attaching a second screen to a running agent. Tradeoff: there's an open-core seam stated in the README — the repo builds do local and LAN access only, while WAN modes ship in the official binaries — and a four-week-old app from a first-time-in-this-category author is early, however good the pedigree.

The purest statement of today's theme is also its thinnest: ccshare shares your live Claude Code session with up to five people via a six-character pairing code, AirDrop-style. Everyone sees the same terminal output and agent conversation in real time; everyone can type. It works with Claude Code, Codex, and other terminal-based agents, over local-network discovery or direct connection. MIT, free, and created yesterday — two stars, launched on Product Hunt the same day the repo went up.

Reach for it for the situations where you currently screen-share a terminal over Zoom and narrate: debugging a stuck session with a colleague, teaching someone how you actually drive an agent, or reviewing a junior's prompting the way you'd review their code. Delete that Zoom call. Tradeoff: everything, stated honestly — this is a day-old, two-star fresh bet whose author says in the open that simultaneous-input semantics and the security model are still being worked out. And the security question is the real one: a live agent session carries your credentials' blast radius, so a join code is a capability grant. Today's memory-heist post on the front page is the demand-side evidence that sessions are full of secrets; a multiplayer session needs scoping the single-player one never did.

Tiptap — the established open-source rich-text editor that a large slice of production web apps already build on — shipped an AI Toolkit that treats the agent as one more participant in the collaborative document. The agent gets tools (tiptapEdit, proofread) to read and edit the document's actual structure — tables, headers, specific text ranges — and its edits land as tracked changes the human accepts or rejects, computed by a new diff engine built for structured documents rather than lines. The part that separates it from chatbot-in-a-sidebar: server-side editing, so an agent can work on the document with no browser open, and the human finds suggestions waiting. Framework-agnostic on the agent side (Vercel AI SDK, LangChain, OpenAI or Anthropic SDK, or your own loop), ships as an npm package, with an alpha "Shorthand" format claiming up to 80% token savings — the author's own number.

Reach for it when your product has a document at the center and your current AI integration is copy-paste from a chat panel. This is the last-click pattern from our 07-02 edition — the agent does everything except the final approval — applied to the document layer, and it's also 07-14's Nobie thesis (one artifact both species operate) shipped by an incumbent with distribution instead of a fresh macOS beta. Delete the copy-from-chatbot-paste-into-editor shuttle. Tradeoff: it's a paid add-on — the code is visible on npm but access is licensed, production-ready-beta, and betting on it means betting on Tiptap's platform, not just its editor.

FlutterFlow launched Campus (204 upvotes, #3 on Product Hunt): a persistent project workspace holding the repo, a terminal, project knowledge, conversations, and agent work in one place, so neither the humans nor the agents rebuild context from scattered Slack threads, tickets, docs, and one-off AI chats. Teammates see each other's work on a shared canvas in real time; agents pick up the same project state. Free, hosted, at campus.flutterflow.io.

The pitch names the actual cost center of agent-heavy teams: context reassembly. Every one-off chat with an agent starts from zero; every human handoff re-explains what the ticket already failed to capture; Campus bets the fix is architectural — one room that accretes the context, with both species reading and writing it. That's adjacent to a watch we've had open since 07-06 (the fragmenting human-side agent surface, waiting for something to unify it), approached from the workspace end rather than the approvals end. Reach for it if your team's real project memory currently lives in whoever was in the Slack thread. Delete some of the re-explaining. Tradeoff: hosted and closed — the room that accretes your project's entire working memory is a room in someone else's building — and this is an app-builder company expanding sideways into collaboration, with free-for-now pricing that will eventually be neither.

boop-agent is an MIT template — its author's word, not a product — for running a personal agent that lives in iMessage. Texts arrive via Sendblue webhooks; a dispatcher reads the conversation history, recalls relevant memories, and either answers or spawns focused sub-agents; replies come back with typing indicators and a draft-confirmation step before anything sends. The notable economics: it runs on the Claude Code or Codex subscription you already pay for, via the Claude Agent SDK or Codex's local app-server runtime — local auth, no API key. macOS for the Apple-data features, Convex for state, a tunnel for the webhook. 1,044 stars, created in April — not a launch; it crossed our trending pool for the first time today.

This is today's theme at its logical end point: the collaboration surface the agent joins isn't a dev tool at all, it's the thread where you talk to actual people. Where 07-06's CodeMote relocated the approval gate to your lock screen, boop relocates the whole agent to your texts. Reach for it as a weekend project chassis — it's a readable TypeScript codebase that's already solved the boring plumbing. Delete the "I'll do that when I'm back at my desk." Tradeoff: the plumbing is a stack of third-party services (Sendblue for the iMessage leg, Convex, ngrok), and the README's own words are the caveat — "not optimized for cost or security... don't trust it with anything you wouldn't trust yourself with." An agent in your texts is an agent one confused webhook away from texting on your behalf.

Off the thread but worth knowing: two hosted launches today sell the management layer above all this — YAGNI (yagni.app, $99/month) packages agents as teams that graduate through Training → Supervised → Autonomous levels based on track record, with your corrections only hardening into rules after three similar edits, and Agently (agently.dev) points a temporal knowledge graph at your whole SaaS stack and runs the back office off it. Both closed, both hosted, both unverifiable beyond their own copy — the org-chart-for-agents thread from 07-07, now with pricing pages. On re-trends: Open Interpreter (65k stars, created 2023) is trending again, its description now reading "a coding agent for low-cost models" — an established project repositioning onto the local-model bet our 06-30 watch flagged — and Clawk, yesterday's footer, added ~120 stars overnight on a 216-point Show HN run; the sandbox shelf note stands.