FIELD NOTES / 30 OCTOBER 2026

What a dry run can—and cannot—tell you.

An AI agent dry run rehearses supported actions without applying their intended changes to production. Its useful output is evidence: a file diff, a proposed destination, a simulated write, or an explicit unknown. A fluent description of what an agent hopes to do is only a plan.

Start with the action, not the confidence score.

Consider an agent asked to “clean up old files.” A useful preview identifies the exact paths, shows deletions, and distinguishes generated output from configuration. If it also proposes uploading diagnostic logs, the review should show the destination and what data would leave. One reassuring score cannot replace those details.

Three layers do different jobs.

A sandbox limits reach. It can restrict which files and network destinations a process can access. Anthropic’s account of Claude Code sandboxing explains why both filesystem and network boundaries matter. Isolation creates a controlled place to work; it does not tell a reviewer whether a proposed change serves the business goal.

A guardrail checks rules. Invariant Guardrails is an existing open-source layer for checking and intercepting LLM or MCP tool calls. Rules can constrain actions, but their coverage depends on the rules and the calls that actually pass through the control.

A rehearsal gathers evidence. It runs supported behavior against an isolated copy, test service, or simulated response. Cloudflare Sandboxes supplies infrastructure for isolated execution; the application still decides which APIs, data, and network access to expose. Designing a trustworthy rehearsal requires those choices.

What deserves a place in the review?

• Changes: exact files or records, with before-and-after evidence where available.
• Boundaries: target environment, requested permissions, and external destinations.
• Side effects: messages, deployments, purchases, or other writes the task proposes.
• Unknowns: unsupported tools, missing context, and results that were mocked.

A rehearsal is not a guarantee.

Production state can change after the preview. A remote API may behave differently from its mock. A retry can duplicate a write. Some actions, including sending a message, cannot reliably be undone. Approval should therefore be tied to the specific reviewed action, with a fresh check when its inputs or target change. Unknown effects should remain visible.

Where AheadTrace fits.

AheadTrace is in development. We are exploring an understandable review layer for developer-agent actions, starting with proposed changes and external side effects. The homepage contains scripted concept examples. It does not run agents, inspect your systems, or provide the protections described above.

Tell us where a preview would help your team. Join the AheadTrace waitlist to take part in product research and hear about early access.