Browser Agent Runtime

A Chrome Manifest V3 extension for conversational, page-aware agents — backed by an explicit runtime (planning, tools, capability detection, policy, validation) and an inspectable Agent Trace.

Talk to the agent about the page you opened it on. Suggestion chips send the same kind of message as typing (no special “mode”). Free-form and chip Goals share one conversational Workflow. Language → Auto detects PreferredLanguage from each message; you can still lock a language manually.

Chrome Built-in AI stays on-device in the side panel.

Session model

Click the toolbar icon on an http(s) tab → a tab-scoped side panel opens for that tab and a new Agent Workspace group (Browser Agent) is seeded with it.

• Switch tabs → the panel hides.
• Open the extension on another tab → a separate Conversation + group.
• Close the panel → session tabs ungroup (they stay open). Reopen = fresh Conversation (no persistence).
• No “Add current tab” step — opening the extension is the session link.

Architecture

Toolbar action on tab T
  → tab-scoped Side panel (?homeTabId=T)     [hidden on other tabs]
  → createSession(T) → new "Browser Agent" group
                     │
                     ├─ Language: Auto or lock → Goal.context.preferredLanguage
                     ├─ Conversation history → Goal.context.conversationHistory
                     ├─ CapabilityRegistry  → strip + Download CTA when needed
                     ├─ AgentRuntime       → Goal → Plan → Workflow → Result (AbortSignal / Stop)
                     │     ├─ Planner (chat / page-grounded / web research)
                     │     └─ WorkflowExecutor
                     ├─ ToolRegistry (page + workspace + Built-in AI)
                     └─ UI: workspace strip · thread · chips · Stop · Trace
  → panel close → endSession → ungroup (tabs remain)

• Planner branches intent inside one conversational Workflow. Research rewrites a concise search query, fetches SERP HTML in the background (DuckDuckGo; Google tab only as fallback), and answers strictly from that evidence.
• Side panel hosts the runtime (Chrome Built-in AI) — one instance per home tab (ADR 0003).
• Expand Runtime Trace under a reply to see Plan / tool Events. Use Stop to abort in-flight work.

Setup

npm install
npm run build

Load unpacked dist/ from chrome://extensions (Developer mode). Dev: npm run dev. Tests: npm test.

Shareable zip (version 1.0.0):

npm run pack

Chrome Built-in AI

Needs Summarizer, Translator, Language Detector, and Prompt on a supported device:

• Built-in AI overview
• Get started

The Capabilities strip shows readiness. Free-form chat needs Prompt available. When models are downloadable, use Download models before sending.

How to use

1. Open an http(s) article or docs page and click the extension icon.
2. If Prompt is downloadable, click Download models and wait until Capabilities show available.
3. Leave Language on Auto, or lock a PreferredLanguage.
4. Type freely or tap a suggestion chip.
5. While Running, use Stop to cancel — or read the reply and expand Runtime Trace.
6. Close the panel to ungroup session tabs (they stay open). Reopen for a fresh Conversation.
7. Switch tabs to hide this panel; open again on another tab for a separate session.

Suggestion	Sent as
Analyze Page	Analyze this page.
Learning Path	Turn this page into a learning path.
Summarize	Summarize this page.
(typed message)	your text

Summarizer/Prompt foundation languages: en, ja, es, de, fr. Other languages (e.g. pt) use Translator after foundation work.

Permissions

sidePanel, scripting, tabGroups, plus host_permissions for http://*/* and https://*/* (side-panel clicks do not get activeTab). Restricted URLs (chrome://, Web Store, etc.) cannot seed a workspace.

What’s next

A Job Analysis Agent should consume this runtime (Conversation + workspace + tools + Trace) rather than reimplementing it.