Checks: Automated Post-Edit Verification

Checks are post-write verification commands (such as typecheckers, linters, and unit test suites) that run automatically after the agent modifies code.

Unlike traditional agents that rely on slow manual testing, Agento automatically runs fast compiler and linter commands after each edit. If a change introduces a syntax error, type mismatch, or failing test, the diagnostics are fed directly into the model’s prompt so it can self-correct immediately.

The Two-Tier Verification Model

To ensure verification is both instantaneous during editing and exhaustive before completing tasks, Agento partitions checks into two tiers:

┌────────────────────────────────────────────────────────┐
│  Tier 1: Fast Micro-Checks                             │
│  - Runs on EVERY file edit                             │
│  - Strict time budget (default: <2000ms)               │
│  - Fast single-file linters, quick typecheck probes    │
└────────────────────────────────────────────────────────┘
                           │ All edits in milestone complete
                           ▼
┌────────────────────────────────────────────────────────┐
│  Tier 2: Milestone Verification                        │
│  - Runs on milestone or task completion                │
│  - Authoritative full test suites, global typecheck    │
│  - Guarantees repository integrity before advancing    │
└────────────────────────────────────────────────────────┘

---

Automatic Discovery

When you launch Agento in a repository, it automatically inspects project configuration files to discover applicable verification commands:

Ecosystem	Detected Files	Auto-Discovered Checks
TypeScript / JavaScript	tsconfig.json, package.json, bun.lock, pnpm-lock.yaml	tsc --noEmit, eslint, biome check
Rust	Cargo.toml, Cargo.lock	cargo check, cargo clippy, cargo test
Go	go.mod	go vet, golangci-lint, go test
Python	pyproject.toml, requirements.txt	mypy, ruff check, pytest

---

Noise Filtering & Regression Attribution

In large codebases, pre-existing compiler warnings or broken tests in unrelated files can overwhelm an AI agent.

Agento uses an attribution engine that isolates only newly introduced regressions:

• Baseline Capture: Captures existing diagnostics before editing.
• Delta Attribution: Only newly created errors originating in touched files are reported to the agent. Pre-existing repo warnings are suppressed from the agent’s turn prompt, preventing context pollution.

---

Custom Configuration (.agento/config.json)

You can define custom checks and override time budgets in your repository’s .agento/config.json:

{
  "$schema": "https://agento.dev/schemas/config.json",
  "checks": [
    {
      "id": "fast-typecheck",
      "command": "bunx tsc --noEmit",
      "tier": 1,
      "timeBudgetMs": 2000,
      "triggerGlobs": ["src/**/*.ts", "src/**/*.tsx"]
    },
    {
      "id": "unit-tests",
      "command": "bun test",
      "tier": 2,
      "timeBudgetMs": 15000,
      "triggerGlobs": ["src/**/*.ts"]
    }
  ]
}

Configuration Options

• id: Unique identifier for the check.
• command: Shell command to execute.
• tier: 1 (runs on every edit within time budget) or 2 (runs during milestone completion).
• timeBudgetMs: Maximum execution time before deferring. If a Tier-1 check exceeds its budget, it is deferred to Tier 2 so file edits never stall.
• triggerGlobs: Array of glob patterns that trigger this check.

---

Inspecting Checks in the TUI

• Press F3 to open the Checks Hub to inspect all registered checks, diagnostic outputs, and execution metrics.
• Press F4 to trigger Sync Checks, running a complete re-scan of the project verification rules.