Share
Introducing pharo-agentic-browser
pharo-agentic-browser is a Pharo-native GUI for orchestrating multiple AI coding agent sessions — Claude Code, Gemini CLI, Cursor CLI, and any ACP-compatible agent — in parallel, right from inside your Pharo image.
No more juggling terminal windows: create a "topic" per task, chat with the agent, and reference live code directly with @ClassName mentions, drag-and-drop, or screenshots — no copy-paste needed. When the AI needs your approval, it pauses right in the conversation; when you set a goal, it keeps retrying until the goal is achieved and reports back.
This slide walks through the core workflow and basic features: supported agents, code mentions, screen captures, goal setting, review by `crit`, and session persistence across image restarts.
Masashi Umezawa
Transcript
37 slides
-
Introducing pharo-agenticbrowser Multi-Agent Session Manager for Pharo Smalltalk Masashi Umezawa
https://github.com/mumez/pharo-agentic-browser -
pharo-agentic-browser An Integrated AI coding tool that allows you to
use multiple AI coding agents — Claude Code, Codex, OpenCode, and others — from within Pharo. 3 -
Core Workflow Each coding session is managed as a topic:
1. Create a topic and select an ACP-compatible agent 2. Type a request + mention code with @ClassName or attach a screenshot 3. The AI works autonomously (task decomposition, code changes, tests) 4. When the AI needs approval, it pauses and asks in the chat 5. Respond with pulldown selection — the AI resumes 6. Topic status is always visible in the sidebar 5 -
AI GUI Tools Are Going Mainstream Dedicated GUI tools for
AI coding agents are becoming standard: Claude Desktop — Claude with tools, MCP, and file access Codex App — OpenAI's autonomous coding environment Cursor / Antigravity / Kiro — AI-native editors These tools lower the barrier for interacting with AI agents beyond simple chat. 7 -
The Shift: IDE → Multi-Agent Delegation Era Paradigm Examples Early
Days Chat in editor sidebar ChatGPT, GitHub Copilot Little While Ago AI-first IDE with agent mode Cursor, Windsurf Now Delegate full tasks to agents Antigravity 2.0, Codex App The AI can now handle entire feature-level tasks, not just line completions. Tools are evolving: the UI is no longer "editor + chat" — it's session orchestration. 8 -
Why Pharo Needs This Too Pharo developers should be able
to leverage the same paradigm: Rich live environment — classes, methods, and runtime are right there pharo-acp already provides ACP client support for multiple agents Multiple projects can run in parallel in one image A Pharo-native GUI that can control multiple AI agents is the natural next step. 9 -
Key Advantages Over Generic Tools 10 Advantage Details Direct context
passing @ClassName , @Class>>method , screenshot — no copy-paste Agent-agnostic Works with any ACP agent via pharo-acp Pharo-integrated Tests, code change watching use the live image Parallel sessions Run several agents on different topics simultaneously -
Installation — Pharo Side Open a Playground in a Pharo
12+ image and evaluate: Metacello new baseline: 'AgenticBrowser'; repository: 'github://mumez/pharo-agentic-browser:main/src'; load: 'all'. Then open the browser: AgenticBrowser open. 12 -
Installation — Agent Side 13 Any ACP-compatible agent - preset
list: Agent Install Claude Code npm install -g @agentclientprotocol/claude-agent-acp Codex npm install -g @agentclientprotocol/codex-acp Gemini CLI ACP built-in ( gemini --acp ) Copilot CLI ACP built-in ( copilot --acp --stdio ) Cursor CLI ACP built-in ( agent acp ) Also supported: OpenCode, Kilo Code, Kiro CLI and more Recommended: Install smalltalk-dev-plugin in your agent for better Smalltalk -
Creating a Topic 1. Click + New Topic 2. Enter
a title and select an agent 3. (Optional) Set an existing project directory 4. Click Create — topic appears in the left sidebar 5. (Optional) Right-click → Set Target Packages... to configure tracked packages The first message is automatically prefixed with /st-buddy to activate the Smalltalk buddy skill. 15 -
Topic States 16 Each topic follows a clear state machine:
#initial promptSent promptSent #working permissionRequested turnEnded humanResponded #endTurn goalReached #goalAchieved #waitingForHuman -
Topic State Icons Icon State working 17 Meaning AI is
actively running ? waitingForHuman AI needs approval • endTurn Turn completed ✓ goalAchieved Goal reached -
Human-in-the-Loop When the AI requests permission: Status changes to ?
(waitingForHuman) Send button becomes Allow Cancel button becomes Deny Respond by clicking the button — the AI resumes seamlessly. No modal dialogs. Approval is part of the conversation flow. 18 -
Code Mentions Reference Pharo classes or methods directly in the
chat: @QueryClass @DBAdapter>>connect please refactor this AgenticBrowser resolves each @mention to its Tonel source and attaches it as an ACP text resource — no copy-paste needed. Drag and Drop Drag a class from the System Browser → inserts @ClassName Drag a method from the System Browser → inserts @ClassName>>methodName 19 -
Screen Captures Click the [ ] button to attach a
screenshot: 1. Click the button — cursor becomes a crosshair 2. Drag to select the area 3. A mention like @sc-20260528-001.png is inserted in the input 4. Send — the PNG is attached as an image resource to the prompt Files are saved to <agenticBrowserRoot>/screenshots/sc-YYYYMMDD-NNN.png . 20 -
File Attachments Click the + button to attach any file
from disk: 1. Click the button — a file selection dialog opens 2. Choose a file — a mention like [filename] is inserted in the input 3. Send — the file's contents are attached as a text resource Oversized files are truncated to maxAttachmentSize (default 5MB) before being embedded in the prompt. Removing the [filename] mention text before sending cancels the attachment. 21 -
Goal Setting Right-click a topic → Set Goal... to enter
a completion condition: all tests pass AgenticBrowser sends a goal prompt to the AI: Goal has been set: all tests pass. When the goal is achieved, summarize and report in result-<topic-id>.md. Keep retrying until the goal is achieved. When result-<topic-id>.md is created, the topic transitions to ✓ ( #goalAchieved ). 22 -
Goal Achievement Hooks Two hooks fire when a goal is
reached: Announcement topic announcer when: AbTopicGoalAchieved do: [:ann | Transcript crShow: ann topic title , ' achieved: ' , ann goal result]. Callback Block topic whenGoalAchieved: [:goal | Transcript crShow: goal result]. Integrate goal events into your own automation workflows. 23 -
Session Persistence Topics are automatically saved to ab-topics.fuel using Pharo's
Fuel serializer — they survive image restarts. Manual save/restore: AbTopicManager save. AbTopicManager load. Per-topic state (settings, status, conversation) is fully persisted. 24 -
Image Change Watching AgenticBrowser watches the image for edits to
a topic's tracked packages: Right-click a topic → Set Target Packages... to configure prefixes (e.g. # ('AgenticBrowser-*') ) When a matching class/method is saved, a system message appears in chat and the package is exported after confirmation Edits to untracked packages are collected — promote them via Apply Updated External Packages Keeps the AI's view of the Tonel source in sync with whatever you change live in the image. 25 -
Code Review with crit crit is a local tool for
reviewing an AI agent's changes in the web browser: Comment on specific lines — the agent receives the comments and replies to them AgenticBrowser starts crit for you and feeds the review back to the topic Setup 1. Install crit and make sure it is on your PATH (e.g. brew install crit ) 2. Check Use crit review in the Settings... dialog The topic's working directory must be a git repository. 26 -
Starting a Review 1. When the agent's turn has ended,
right-click the topic → Review by crit 2. crit opens in the web browser — add comments to changed lines 3. Press Finish Review — the review is sent to the agent 4. Comment again, or approve when satisfied 27 -
Comments and Replies in crit The agent replies in place,
and its changes appear as a new review round. 28 -
Topic Template Every new topic's working directory is seeded from
<agenticBrowserRoot>/topictemplate : Ships with a default CLAUDE.md / AGENTS.md tuned for the smalltalk-dev-plugin Drop in .claude , .opencode , or other agent config directories — skills, commands, rules shared across all topics Replace the defaults with your own customized versions any time Saves you from reconfiguring the coding agent for every new topic. Not applied when a topic points at an existing project directory. 30 -
Adding MCP Servers Place mcp.json in your AgenticBrowser root directory:
{ "mcpServers": { "my-server": { "command": "uvx", "args": ["my-server-package"], "env": {"API_KEY": "value"} } } } Built-in smalltalk-interop and smalltalk-validator servers are auto-merged by default Set useDefaultMcpServers: false to use only your own mcp.json 31 -
Adding Custom Agents From a Playground AbSettings default codingAgents: (AbSettings
default codingAgents copyWith: {'name' -> 'my-agent'. 'command' -> #('my-agent' '--acp')} asDictionary). AbSettings save. Or edit ab-settings.json directly The agent appears as a preset in the New Topic dialog on next open. 32 -
Global Settings 33 Key Default Description useDefaultMcpServers true Merge built-in
Smalltalk MCP servers aiPermissionWaitTimeoutSeconds 1800 Timeout for human approval #reject_once Auto-response: allow_once , allow_always , reject_once 30 Timeout for package export approval aiPermissionTimeoutOption exportApprovalWaitTimeoutSeconds Settings can also be configured per-topic via right-click → Edit Settings... -
Other Interfaces: Web UI & Scripting API Two more ways
to work with AgenticBrowser, beyond the Spec UI: Web UI — web browser interface over WebSocket; mobile friendly, live topic updates → Web UI slides Scripting API — simple DSL to build and run multi-agent orchestrations from code (sequential / parallel steps, save & load) → Scripting API slides 34 -
Summary pharo-agentic-browser brings the multi-agent delegation paradigm to Pharo: Native
GUI — manage multiple AI sessions without leaving your image Agent-agnostic — any ACP-compatible agent works Rich context — code mentions, drag-and-drop, screen captures Human-in-the-loop — conversational approval, no interruption dialogs Goal-driven — set completion conditions, hook into results Browser-based review — comment on the agent's changes with crit; it replies and iterates Extensible — custom MCP servers and agents with minimal configuration 36