[harness] Add Cloud Agent API providers (Cursor, Copilot, Claude Agent SDK)

[santoshkumarradha]

Parent Epic

Part of #291 — Add Cursor as a Harness Provider (Cursor Cloud component)

Summary

Beyond local CLI harnesses, several providers now offer cloud-hosted agent APIs that can execute tasks remotely. This issue tracks research and implementation of cloud agent providers as a new category of harness.

Cloud Agent API Landscape (as of March 2026)

Provider	API	Maturity	How It Works
Cursor	REST API at api.cursor.com/v0/agents	Beta (all plans)	POST to launch agent, poll/webhook for status, operates on GitHub repos, creates PRs
GitHub Copilot	REST + GraphQL API	Production (Pro/Business/Enterprise)	Assign issues to Copilot via API, runs as background agent, creates PRs
Anthropic Claude	Claude Agent SDK (claude-agent-sdk)	Available (Python + TypeScript)	Programmatic query() function with streaming, local execution, not a hosted API
OpenAI Codex	ChatGPT UI only	No public REST API	Cloud agent in ChatGPT sidebar, no programmatic access yet

Design Considerations

These are fundamentally different from CLI providers:

Aspect	CLI Providers	Cloud Agent Providers
Execution	Local subprocess	Remote HTTP/REST
File access	Local filesystem	GitHub repos (or none)
Output	stdout (text/JSON/NDJSON)	REST polling or webhooks
Latency	Real-time streaming	Async (minutes to hours)
State	Session ID for resume	Agent ID for follow-up
Cost model	Per-token (user's API key)	Per-task or subscription

Open Questions

1. Should cloud agents implement the same Provider interface? They're async by nature — Execute() would need to either block-and-poll or return a handle for later collection.
2. Should this be a separate CloudProvider interface? With methods like Launch(), Status(), Collect(), Followup(), Stop().
3. How to handle file I/O? CLI providers read/write local files. Cloud agents operate on remote repos. The .agentfield_output.json pattern doesn't directly apply.
4. Webhook vs polling? Cursor supports webhooks. Copilot uses status polling. How to unify?

Suggested Approach

Phase 1: Implement Cursor Cloud as the first cloud provider — it has the most complete REST API.
Phase 2: Add Copilot (if GitHub API access is available).
Phase 3: Evaluate Claude Agent SDK (it's more of a library than a cloud API).

Cursor Cloud API Reference

POST   /v0/agents              — Launch agent (prompt, model, repo, branch, webhook)
GET    /v0/agents              — List agents
GET    /v0/agents/:id          — Agent status (CREATING → RUNNING → FINISHED)
GET    /v0/agents/:id/conversation — Conversation history
GET    /v0/agents/:id/artifacts    — Generated files
POST   /v0/agents/:id/followup    — Follow-up prompt
POST   /v0/agents/:id/stop        — Stop agent
DELETE /v0/agents/:id              — Delete agent
GET    /v0/models              — Available models
Auth:  Basic auth (API key as username)

Acceptance Criteria

• Design decision on interface (unified vs. separate CloudProvider)
• Cursor Cloud provider implementation (Python SDK first)
• Copilot provider implementation (if feasible)
• Async execution model (launch → poll/webhook → collect)
• Tests with mocked HTTP responses
• Documentation for cloud vs. local provider differences