💡 Cloud-Persistent Agent Sessions: Background Workflow Execution Beyond App Lifetime

[github-actions[bot]]

Summary

Leverage Claude Managed Agents (public beta since April 2026) to run long-running workflows in the cloud, continuing execution even after the user closes TalkTerm. The avatar sends an OS notification when the workflow completes and presents a full briefing with results when the user returns. This addresses Claude Cowork's key limitation (workflows stop when the app closes) while leveraging TalkTerm's BYOK model for unlimited background execution without plan-tier gatekeeping.

Market Signal

Claude Cowork launched cloud remote sessions in July 2026 but gates them to the $200/month Max plan — the highest tier. ChatGPT Work supports background agents but requires a paid subscription. The market expectation is shifting toward workflows that survive app closure — users expect to start a competitive analysis, close their laptop, and get results later.

Claude Managed Agents reached public beta in April 2026 with the managed-agents-2026-04-01 header version. It provides hosted stateful agent execution with session persistence, cost/turn limits, and memory (also in public beta). TalkTerm's BYOK model means background execution costs come from the user's own API key — no plan-tier gatekeeping, no 5-hour rolling windows.

The June 2026 billing change introduced dedicated monthly credits for non-interactive use (GitHub Actions, headless agents, Agent SDK) that don't roll over, making cost predictability for background agent execution even more important.

User Signal

Non-technical users (TalkTerm's primary audience) often start workflows that take 15-30 minutes — competitive analysis across multiple sources, multi-document brainstorming sessions, research reports. They shouldn't have to keep TalkTerm open and their computer awake the entire time. Existing ideas #232 (Goal Mode: Checkpoint-Based Background Workflow) and #247 (Asynchronous Task Review Queue) identify this need from different angles but predate the Managed Agents API availability.

Technical Opportunity

TalkTerm's AgentBackend abstraction layer (defined in src/shared/types/ports/) is designed for exactly this kind of backend swap. A ManagedAgentBackend implementation could delegate long-running workflows to the cloud while a LocalAgentBackend runs in-process for interactive sessions.

Implementation architecture:

1. User starts a workflow and chooses "run in background" (or the avatar suggests it for long workflows)
2. ManagedAgentBackend dispatches to Claude Managed Agents API with the user's API key
3. App can be closed; a lightweight polling service checks for completion
4. On completion, Electron's Notification API delivers an OS notification
5. When the user returns, the avatar presents a briefing: "While you were away, I completed the competitive analysis. Here's what I found."
6. The DomainEventBus emits session:completed-remotely to trigger the briefing flow

The existing cost/turn limits (maxTurns, maxBudgetUsd) from the Claude Agent SDK apply to managed sessions too, preventing runaway costs.

Assessment

Dimension	Score	Rationale
Feasibility	med	Managed Agents API is public beta (3+ months stable) but not GA yet
Impact	high	Solves real user pain; differentiates against Cowork's Max-plan gating
Urgency	med	Cowork just shipped cloud remote; TalkTerm has a window to match with better economics

Adversarial Review

Strongest objection: Building on a beta API (Managed Agents) risks breaking changes. Background execution with the user's API key unattended creates runaway cost risk.

Rebuttal: The Managed Agents API has been stable for 3+ months with a versioned header (managed-agents-2026-04-01), suggesting it's approaching GA. Implementation should be opt-in with clear cost estimates before dispatching to cloud — the avatar says "This analysis will cost approximately $2.50 and take about 20 minutes. Should I run it in the background?" The SDK's maxBudgetUsd provides a hard cost ceiling. If the API changes, graceful fallback to local-only with a user notification.

Suggested Next Step

Evaluate the Managed Agents API stability, cost structure, and session persistence model. Design the ManagedAgentBackend implementation of the AgentBackend interface. Prototype with a simple workflow that dispatches to cloud, polls for completion, and triggers an OS notification.

---

Sources: Claude Managed Agents (Tygart Media), Claude Cowork Review (Intelligence & Power), Claude Agent SDK Guide (Totalum)

[github-actions[bot]]

Weekly Update

What Changed

Anthropic launched the Managed Agents API — a cloud runtime that directly enables the cloud-persistent sessions envisioned in this proposal:

• Three new API resource types: /v1/agents (agent definitions), /v1/sessions (stateful conversation threads), /v1/environments (sandboxed execution containers).
• Cloud containers run agents in stateful environments billed on session-hours, not just token consumption.
• Built-in scheduler and dreaming pass — the runtime supports periodic agent wake-ups and self-improvement cycles between sessions.
• Rubric-based outcome grading enables quality gates on autonomous work.

This eliminates the need for custom cloud infrastructure proposed in the original idea. Anthropic's native runtime provides session persistence, background execution, and state management out of the box.

Additional context: Claude Fable 5 (Mythos-class, $10/$50 per M tokens) launched July 2026, specifically designed for the long-running autonomous patterns this proposal envisions. Combined with Managed Agents, TalkTerm could offer "Deep Work" sessions that run in the cloud beyond app lifetime with periodic avatar-narrated check-ins.

Sources: Claude Agent SDK Documentation, Anthropic Managed Agents

Updated Assessment

Dimension	Previous	Current	Delta
Feasibility	med	high	↑ Managed Agents API provides native cloud runtime — no custom infrastructure
Impact	high	high	→ Background execution beyond app lifetime remains high-value
Urgency	low	med	↑ API is available; Fable 5 provides the ideal model for long-running sessions

Recommendation

Advance. The Managed Agents API transforms this from a custom-infrastructure project into an API integration. Recommend re-scoping: (1) Use Managed Agents /v1/sessions for cloud persistence instead of custom backend, (2) Combine with Fable 5 for long-running autonomous workflows (see new proposal #434), (3) Use the built-in scheduler for periodic check-ins rather than building custom notification logic. The heavy engineering is now handled by Anthropic's runtime.