Skip to content

The Two Guardrails

Agento uses a two-tier guardrail system designed to prevent broken code and eliminate hallucination drift during autonomous coding sessions.

Proposed Edit (write_file / edit_file)
┌────────────────────────────────────────────────────────┐
│ LINTOs (PRE-WRITE, blocking) │
│ - Evaluates rules against proposed in-memory buffer │
│ - Hard gate -> Rejects edit immediately │
│ - Justify gate -> Requires explicit human consent │
└──────────────────────────┬─────────────────────────────┘
│ Passed
File Written to Disk
┌────────────────────────────────────────────────────────┐
│ CHECKS (POST-WRITE, non-blocking feedback) │
│ - Runs fast diagnostics (compiler, typecheck, lint) │
│ - Extracts newly introduced errors │
│ - Surfaces actionable feedback to agent prompt │
└────────────────────────────────────────────────────────┘

LINTOs are deterministic rules that evaluate proposed code changes before they are written to disk.

  • Execution Timing: In-memory, immediately prior to file mutation.
  • Behavior: Blocking. If a rule is violated, the file write is stopped.
  • Purpose: Enforcing non-negotiable repository standards (e.g. banning any casts, preventing hardcoded credentials, protecting configuration files, and ensuring architectural boundaries).
  • Gate Levels:
    • hard: The mutation is rejected immediately, returning rule feedback to the model to rewrite the code.
    • justify: Pauses execution to require human approval or a justified override.

Checks are non-blocking diagnostic commands (such as tsc, cargo check, eslint, or unit tests) that run after an edit is successfully written to disk.

  • Execution Timing: Immediately after a file write or on milestone completion.
  • Behavior: Non-blocking feedback. Checks never prevent a file from being written; instead, they capture compiler and linter output to provide the agent with instant, real-world feedback.
  • Noise Filtering: Agento filters out pre-existing repository errors and reports only the new errors introduced by the agent’s recent edits.
  • Tiers:
    • Tier 1 (Fast Micro-Checks): Rapid checks running within strict time budgets (<2s) on every edit.
    • Tier 2 (Milestone Verification): Authoritative test suite and full build checks run upon completing a milestone.