Agent Planning-Mode (Claude Agent SDK + e2b)
An AI-cofounder / autonomous-startup-builder SaaS
Overview
An early build for an AI-cofounder / autonomous-startup-builder SaaS exploring “planning mode” with the Claude Agent SDK and an e2b code-interpreter sandbox, a scratch build that fed into the production sandbox and agent-base designs.
Why It Exists
Before committing to a sandbox runtime and agent-control protocol, the team needed to validate how an agent plans and executes against an isolated code-interpreter and how to persist run state. This build is that exploration, kept honest as throwaway R&D rather than production code.
What We Built
A small TypeScript spike: planning-mode-build.ts driving the @anthropic-ai/claude-agent-sdk against @e2b/code-interpreter, with better-sqlite3 for local run/state persistence and ws for a WebSocket channel. A single plan script runs it. The repo is intentionally minimal (a planning script, an e2b/ folder and a debug log), reflecting its production nature.
Technologies & Approach
TypeScript with the Claude Agent SDK and e2b, the same building blocks later hardened in the production sandbox bridge. SQLite kept iteration fast and local while testing the plan/execute loop.
Outcome / Impact
Validated the agent planning-and-execution loop against a sandboxed interpreter and informed the architecture of the production e2b sandbox, bridge protocol and agent-base repos that followed.
Capabilities Demonstrated
- Rapid building of an agent planning loop
- Claude Agent SDK driving sandboxed code execution (e2b)
- Local state persistence for agent runs (SQLite)
- De-risking architecture decisions before production build-out