Project Think roadmap: durable agent platform preview to stable developer experience

[threepointone]

Project Think shipped the first preview of the next-generation Agents SDK harness: durable chat execution, Session-backed memory, sub-agent orchestration, workspace tools, codemode execution, extensions, and the beginnings of a full execution ladder.

The initial Think parity roadmap is mostly complete. This umbrella tracks the next phase: turning the preview into a coherent, stable developer experience, with clear docs, examples, integration stories, and focused hardening.

This issue is intentionally a hub. Child issues should be small and independently reviewable.

Product goal

Make Think the default path for building durable, serverless agents that can think, act, persist, fork, and hand work off across Cloudflare infrastructure.

Success means a new contributor can answer, from this issue and its children:

• What has already shipped?
• What blocks a stable preview / eventual 1.0-quality experience?
• Which issue or PR owns each remaining piece of work?

Track 1: Core Think reliability and turn semantics

• apply recovery to Think.chat() #1429: Apply recovery to Think.chat() so sub-agent/RPC chat turns get the same durable recovery story as normal chat turns.
• Chained turns: pattern + helper for multi-phase auto-continuation #1386: Design chained turns / multi-phase continuation for long coding and research tasks without persisting synthetic user messages.
• Think + external LLM providers broken: Zod v4 tool schemas produce empty JSON Schema #1322: Re-triage external provider + Zod schema compatibility against current dependency versions.
• think: lifecycle hook follow-ups (from #1340) #1343: Finish lifecycle hook polish/tests. Treat as hardening, not a launch blocker.
• Document chatRecovery / onChatRecovery consistently across Think docs, durable execution docs, and server-driven message docs.

Track 2: Multi-session and app/product shape

Current direction: one Think child Durable Object per conversation, with a parent Durable Object owning directory/sidebar state and shared user-level resources. This matches the examples/assistant direction and answers most of #1349.

• Write docs/think/multi-chat.md for the parent/child multi-chat pattern.
• Decide whether useChats() should move from example-local code into agents / agents/react.
• Decide whether a Chats base class is worth promoting, or whether the pattern should stay documented with examples for now.
• Document the shared-resource boundary from examples/assistant: shared workspace + shared MCP, but per-chat messages, memory, extensions, and branch history.
• Measure / document parent DO scale limits for shared workspace and shared MCP fan-out.
• Decide whether Think: Is "one agent, many concurrent chat sessions" supposed to work as a product model in Think? #1349 can be closed after the docs and example guidance are clear.

Track 3: React/client package boundary

Think currently relies on @cloudflare/ai-chat/react for shared chat UI behavior, even though much of the protocol/runtime substrate has moved into agents/chat.

• Hoist shared React chat primitives from @cloudflare/ai-chat/react into agents or agents/react, with compatibility re-exports.
• Address Think-native initial message behavior so Think can avoid the redundant HTTP get-messages fetch / Suspense flash when WebSocket history is already authoritative.
• Link and triage related client backlog: Support non-suspending hook + exportable fetch function for framework loaders #1011, Support fallbackMessages for instant conversation switching #1045, useAgentChat can throw "Maximum update depth exceeded" during dense streamed tool-input-delta bursts #1361, Replace client-defined tool round-trips with server-side RPC delegation #1414, Support client-side tools in headless AIChatAgent agent-tool runs #1420.
• Keep Replace client-defined tool round-trips with server-side RPC delegation #1414 as discussion for now. Client-tool RPC delegation is interesting but likely too invasive for the immediate roadmap.

Track 4: Execution ladder and tools

Think should remain useful at Tier 0 with only the workspace tools. Each higher tier should be additive and explicitly configured.

• Land/review Add multimodal workspace read support #1435 as the concrete fix for Think: workspace read should support multimodal file/image content #1403 multimodal workspace reads.
• Decide whether Memory: support multimodal content (files, images, audio, artifacts) #1392 multimodal memory is a separate Session/memory track after Add multimodal workspace read support #1435 lands.
• think: make createSandboxTools actually work #1319: Make createSandboxTools actually work, or clearly mark the export as a placeholder until Sandbox integration is ready.
• think: add a fetch tool with whitelisted domains #1344: Design a whitelisted fetch tool that fits Think's capability model.
• Track browser/tooling follow-ups such as browser: allow for tool executions on a single instance #1398 and browser side rpc #1401.
• Track MCP ergonomics and compact-tool-surface work: addMcpServer({ autoRegister: false }) — let consumers suppress MCP tool auto-merge to enable compact / search+execute patterns #1336, addMcpServer: default callback URL silently breaks when callers don't set callbackPath + sendIdentityOnConnect: true #1378, MCP streamable-HTTP transport: large JSON-RPC payloads (>16 KB) fail with TLS record_overflow due to header-encoding #1433, Fix large MCP streamable HTTP payloads #1434.

Track 4a: Think + Artifacts integration

Artifacts looks like the missing versioned handoff layer for Think workspaces and sessions: one repo per agent/session/task, forkable history, Git-compatible tooling, and short-lived repo-scoped tokens.

Start with focused child issue #1440: Think + Artifacts: versioned workspaces, forks, and handoff.

Track 5: Examples, apps, and docs polish

Think needs one clear reference path for developers, plus smaller docs/examples that explain each capability without requiring readers to reverse-engineer the kitchen-sink app.

• Use examples/assistant as the primary Think reference app and document which pieces are canonical versus app-specific.
• Decide what to do with draft PR feat: think-cli + think-server — remote coding agent #1135 (think-cli / think-server): revive as an app track, split into smaller issues, or close/supersede.
• Add / update Think docs for missing multi-chat guidance, peer dependency wording, Think vs AIChatAgent, hooks, client tools, recovery, and execution-ladder limitations.
• Keep examples/multi-ai-chat positioned as a lower-level Agents/chat example unless it is intentionally upgraded to Think.

Track 6: Documentation and design hygiene

Several design and WIP docs still describe pre-Session or pre-shared-chat-layer assumptions. This is not all launch-blocking, but stale docs make it hard for humans and agents to pick up roadmap work safely.

• Refresh canonical Think design docs: design/think.md, design/think-roadmap.md, and design/think-sessions.md.
• Update shared chat design docs whose status now conflicts with shipped code: design/chat-shared-layer.md, design/chat-improvements.md, and design/think-vs-aichat.md.
• Add Think cross-links from adjacent docs: sessions.md, durable-execution.md, long-running-agents.md, server-driven-messages.md, workspace.md, codemode.md, browse-the-web.md, mcp-client.md, client-sdk.md, workflows.md, and observability.md.
• Fold still-relevant WIP notes into permanent docs/design, especially wip/think-multi-session-assistant-plan.md and wip/inline-sub-agent-events.md; mark the rest historical or remove when safe.

Track 7: Experimental learnings and promotion candidates

The experimental folder has useful prior art, but it should inform Think deliberately instead of becoming hidden roadmap scope.

• Evaluate experimental/session-skills as a possible Think + Session skills reference app.
• Decide whether Session experiments (session-memory, session-search, session-multichat) need Think-first variants or should remain lower-level Agent + Session examples.
• Pull durability lessons from forever-chat, forever-fibers, and inference-buffer into the Think recovery track without making AI Gateway buffering a Think blocker.
• Use gadgets-* experiments as background for facets/sub-agent safety, without blocking Think’s stable preview on experimental loader/gatekeeper work.
• Update experimental/README.md so useful prior art is discoverable.

Flagship Project Ideas

These are aspirational apps that should guide Think’s product direction. They are not stability blockers, but they should pressure-test the SDK, docs, examples, and integration story.

• Personal assistant / life OS: A long-lived assistant modeled after OpenClaw, with memory, calendar/email/tasks, browser use, documents, reminders, and human approval for sensitive actions.
• Coding agent: A durable coding agent with workspace, Artifacts/Git, Sandbox, Browser Run, tests, PR creation, code review, and recoverable long-running tasks.
• SMB operating hub: A hub for managing a small business: inbox triage, customer follow-up, invoices, bookings, inventory, support, website updates, marketing drafts, and analytics.
• Research analyst: A research workspace with browser crawling, citations, source snapshots, knowledge base, report generation, and follow-up monitoring.
• Customer support copilot: A persistent assistant per customer/account that knows past tickets, product usage, docs, changelogs, and can draft responses or run diagnostic workflows.
• Agentic browser QA lab: A Browser Run powered QA agent that explores app flows, records sessions, finds regressions, captures screenshots, and files issues.

[threepointone]

Opened the Artifacts child issue: #1440

This covers the versioned workspace/session handoff story: one repo per Think session/task, forkable chat state, Git notes/metadata, and Sandbox/ArtifactFS handoff. It should start as an additive integration around Think's DO-local Workspace rather than replacing Workspace itself.