A coding agent that interrogates its own changes before making them

What it does

Autoloom is a free Electron-based desktop client (Windows x64 for now) for AI coding, built on DeepSeek’s open-source DeepSeek Harness with the Cordis plugin mechanism intact. You bring your own model and API key. The twist is that it doesn’t just run your agent — it injects engineering-governance checkpoints at decision points: questioning whether a change is needed, checking for unnecessary complexity before editing, and reviewing delivery against requirements and actual execution evidence.

The interesting bit

The governance isn’t a vibe or a system-prompt paragraph — it’s the Aegis method pack (a separate open-source project) wired into the runtime, with prerequisite ordering for key operations and process records retained. Each project mutation can show a collapsed “Governance impact” row explaining how the methods shaped that specific action, and recorded goals/constraints/decisions persist with the project and surface as reviewable diffs when changed.

Key highlights

• Four method families: change-necessity (first principles), anti-entropy/complexity checks, counterexample review and causal diagnosis, and completion review against verification evidence.
• Local projects plus Linux x64 targets over SSH, with code diffs, command logs, previews, and docked interactive terminals.
• Built-in Parallel/Exa MCP web search — no separate search API key.
• Configurable test-first preference and code complexity thresholds.
• Session backup, background Beta updates, and SHA256 checksums per release.

Caveats

• The client source is closed for now — open-sourcing is “under evaluation,” so this is a product, not a hackable codebase.
• Windows x64 only in Beta; macOS and Linux clients are planned but undated. The installer is unsigned, so expect SmartScreen grumbling.
• The README is candid that governance prompts don’t guarantee reliability — you still judge results from code and verification evidence.

Verdict

Worth a look if you’re tired of agents that refactor first and ask questions never — the evidence-trail angle is genuinely unusual. Skip it if you want an open-source, cross-platform tool today; this is a closed Beta with one platform and a philosophy to sell.