Speculative Execution in Claude Code, Hidden but Fully Built

Summary

A developer reverse-engineered Claude Code’s binary and discovered a fully built but server-gated feature called “speculative execution” - Claude predicts your next action, pre-executes it in a sandboxed overlay filesystem, and delivers instant results when you accept. Combined with the recently shipped “auto mode” permissions, this could eliminate the entire post-coding workflow (tests, docs, commits) by running them speculatively in the background.

Key Insight

• Speculative execution is fully built inside Claude Code but gated behind a server-side flag called “Tengu speculation” that returns false. The overlay filesystem, permission tiers, and pipelining are all implemented.
• How it works: After completing a task, Claude generates a prompt suggestion (e.g., “run the tests”), then immediately starts executing that suggestion in a sandboxed overlay filesystem. File writes go to the sandbox, not the real codebase. If you accept, the overlay syncs back. If you type something else, it gets discarded.
• Permission tiers are baked in: Read-only tools (file reads, grep, ls) run freely during speculation. File edits and shell commands only execute if previously approved. Speculation halts at any unapproved permission boundary.
• Chained speculation: When one speculative execution completes, it immediately generates the next prediction and starts executing that too - staying multiple steps ahead.
• Auto mode + speculation = the real play: The recently shipped “auto mode” (intelligent command classifier) removes the permission wall that currently makes speculation impractical. Together, they’d automate the entire best-practices pipeline (test, lint, commit, document, PR) without the developer ever asking.
• Anthropic is tracking metrics: Time-save measurements and speculation acceptance rates are visible in the binary, suggesting this is heading toward public release.