Skip to main content
This page will take you through the mechanism Maestro uses to actually complete engineering tasks, ie, how files are created, modified and validated. You may be familiar with systems that modify files using tools, which are immediately executed upon invocation. This style of agentic engineering can lead to sloppy, poorly planned changes, and carrying out large scale additions or refactors across multiple files can be inefficient and fragmented. A system for real world engineering needs to be able to freely think as it works, allowing for spontaneous backtracking or appending of additional changes when necessary. Maestro accomplishes this using a unique Propose then Apply system.
1

Propose Changes

Maestro will propose changes to files, interweaving proposals with tool calls to solve real engineering tasks.
2

Apply Changes

When ready, Maestro explicitly calls the Apply Changes tool, which intelligently completes all proposals. Multiple validations on the correctness of the code are applied on a batched basis at this point.
3

Oversee Changes

The final diffs will be evaluated against the original proposal intents, rejecting anything with issues. This leverages multiple adversarial systems to cross validate code quality.

​
Proposing Changes

Maestro can interleave proposals with tool calls to gather additional context, append new changes to the proposal stack if it uncovers additional work, or backtrack if it notices an issue with previously proposed code. This allows for extensive and flexible engineering workflows. Example Proposal
Stopping a running session with the stop button will discard any unapplied proposals. Maestro will remain aware of these discarded proposals if you wish to ask to recover them.

​
Applying Changes

Once Maestro decides that the prior proposals represent a correct set of changes, the Apply Changes tool is used to attempt to implement the proposals. This system is flexible. You can steer Maestro towards small, incremental changes, or multi-file refactors. Once this tool completes, all proposals will either be in an Accepted or Rejected state. We want to reward good progress and keep moving forward; concurrent sets of changes may be partially accepted, i.e, if 3/5 files were approved, new iterations will be created for 3 files, and no change implemented for the other 2.

​
User Approval

Example Apply When Maestro applies changes, you’ll see a confirmation dialog showing:
  • File diffs: Each proposed change is displayed with clear before/after comparisons
  • Change summary: A description of what’s being modified and why
  • Approval options: You can review each change individually
The approval flow provides several key benefits:
  • Safety checkpoint: Review changes before they’re committed to ensure they match your expectations
  • Partial acceptance: Accept some changes while rejecting others if needed
  • Learning opportunity: Maestro learns from rejected changes to improve future proposals
Auto-accept mode: Selecting auto-accept will automatically approve all changes until the end of the turn, streamlining the workflow for trusted operations. Use this when:
  • Working with well-tested code patterns
  • Making repetitive similar changes
  • You have high confidence in Maestro’s understanding of the task
When to manually review: Keep manual review for:
  • Critical system components
  • Security-sensitive code
  • Complex architectural changes
  • First-time implementations of new patterns
Sometimes during application of changes, you may see a message regarding anti-hallucination, preventing the apply operation. This is a safety mechanism to prevent Maestro from making changes to files it has not yet seen. The system will recover automatically, there is no need to do anything.

​
Overseeing Changes

Example Oversee The Overseer system provides an additional layer of quality control:
  • Intent validation: Ensures proposed changes match the stated intent
  • Code quality checks: Validates syntax, structure, and best practices
  • Rejection feedback: Provides specific guidance on what needs to be fixed
  • Iterative improvement: Maestro can revise and resubmit based on feedback
This multi-stage review process ensures that only high-quality, correct changes are applied to your codebase.