Making the agent smarter, more autonomous, and more fun

Underneath, the session sync engine and V2 session core are rolling out. They're the plumbing that makes everything below possible.

• Multi-agent pipelines: one agent plans, one codes, one reviews
• Automatic agent selection based on prompt intent

• Repo convention learning: the agent picks up your codebase's style and sticks to it
• Compounding memory: every session teaches the next one
• Memory introspection: ask the agent why it believes something and where it learned it
• Cross-session recall: "do it like we did in that auth refactor last month"

• Test-aware refactor loops: change, run, verify, repeat until green
• Confidence scoring: the agent tells you when it is guessing
• bolt review --staged / --branch: one-shot AI diff review with exit codes
• bolt commit: commit messages you don't have to rewrite
• Self-review pass before any diff is handed to you

• Watch mode: tests rerun on every save and failures fix themselves
• bolt cron: scheduled agent chores (dependency bumps, changelog drafts, issue triage)
• On-red-main automation: bisect, blame, and propose the fix before you've seen the alert
• Flaky test detection and quarantining
• Background job queue: run --background plus bolt jobs list/tail/kill

• Guardrail agent that vetoes risky commands before they run
• Budget guards: --max-cost and --max-tokens on any run
• bolt undo: one-command rollback when an experiment goes sideways
• Named checkpoints: save points you can rewind the repo and the conversation to

• Design-to-code: hand the agent a Figma file, get components back
• Codebase visualization maps drawn by the agent
• Pair-programming mode with a shared cursor
• Session replays you can share like a highlight reel

• Autonomous long-horizon projects spanning days
• Fine-tuned repo-specific models
• On-device small-model routing for trivial tasks
• Agent federation across organizations
• AI release manager: cut, verify, and publish releases end to end

• Best-of-N runs: the same task fanned out to several models in parallel, keep the winner
• Push-to-talk voice input in the TUI
• Agent evaluation benchmark harness (bolt eval)
• Project memory module: automatic capture with save and recall tools
• bolt memory: view and edit what the agent remembers per project (/memory in the TUI)
• Per-session memory controls
• /plan mode, /todos task list, and /btw side questions that don't interrupt the task
• bolt logs with tail and follow
• .boltignore-aware agent file search
• macOS desktop app with a one-line terminal installer (no Gatekeeper detour)