cooleel

Breakpoint — a debugger for AI agents. Every tool call the agent makes is checkpointed inside a Tensorlake sandbox (filesystem + memory + processes). Opus 4.7 reads the full trajectory — reasoning, tool I/O, filesystem deltas — and points at the causal tool call that broke the run. You can rewind to any checkpoint, scrub through the timeline, live-shell into past sandbox state, or fork a fresh agent from any step with a corrective prompt. Drop-in for Claude Agent SDK; framework-agnostic adapter (bash / edit_file / view) works with OpenAI Agents SDK and any harness that registers Python functions as tools. Problem it solves Agent failures are causal, not local. A model wipes a database on turn 3, looks fine, succeeds on turns 4–9, then fails an integration test on turn 10 with a confusing error. Existing tracers (LangSmith, Langfuse) show you the crash site — turn 10. The actual fix lives at turn 3. That's a git bisect problem on a single agent run, and today there's no good way to do it: - Logs and traces show messages, not state. You can read what the agent said, but not what the filesystem, processes, or database looked like when it said it. The real evidence — a stack trace in /tmp/server.log, two missing rows in todos.db, a half-written config — is gone by the time the run ends. - Re-running doesn't reproduce. Agents are non-deterministic; reproducing a specific failure path can take many tries. - There's nowhere to fork from. Even if you spot the bad step, you can't restart the agent from that exact state with corrective guidance — you have to re-run from scratch and hope it makes the same mistakes up to that point. Breakpoint fixes all three: every tool call is a rewindable checkpoint of full sandbox state, Opus 4.7 does the bisect for you (and can cat log files in any past snapshot to confirm), and any checkpoint is a fork point — restart from there with a fix, watch it run live, get an automatic verdict on whether the fork actually succeeded.