Evidence-First Engineering: Keeping Agent Builds Honest

Where this fits

You ask an agent to take over a half-finished feature in your codebase. It responds “Implemented and verified.” But when you inspect the branch, there are no test logs, no scope notes, and the only “verification” is a passing local run you can't reproduce. Evidence-first engineering is a foundation skill that forces the agent to separate what it has proven from what it assumes, so handoffs, status reports, and release decisions rest on artifacts you can inspect.

Why agents benefit

• Agents can tag each claim as [evidence:path:line] or [assumption], making status reports auditable rather than persuasive.
• Stage gates prevent an agent from jumping ahead to implementation before it has discovered facts and defined the change boundary.
• Sensitive-information redaction keeps generated summaries and public issue posts free of tokens, paths, and internal details.
• Explicit confirmation rules force a pause before risky or external actions, such as writing to a database or opening a PR.
• The skill's resumable continuity model lets an agent recover gracefully after an interruption by reading persisted evidence rather than starting over.

Practical scenarios

Handing off a half-built feature

An agent joins a project mid-sprint and must inventory what already works. Instead of trusting inline comments, it collects build artifacts, existing test results, and git history as evidence before proposing the next step.

Deciding release readiness

A developer asks, “Can we ship this?” The agent produces a readiness report where each checklist item points to a test run, a review comment, or an explicit assumption. Unverified claims are labelled as such, so the developer knows exactly what remains risky.

Recovering from an interrupted run

A long multi-stage task—say, migrating a payment module—hits a timeout or a network failure. On restart, the agent reads the saved evidence file, figures out which stages passed, and resumes from the first incomplete gate instead of repeating or silently skipping work.

Add it to your agent workflow

Install it with the skills CLI:

npx skills add NocodeMrLi/mini-program-engineering-skill-suite --skill evidence-first-engineering

Then tell your agent to apply the discipline explicitly:

For this task, follow evidence-first engineering. Before implementing, list the facts you have verified and state open assumptions. After each stage, store evidence artifacts and mark the stage as done only if the evidence matches the status model.

The skill is domain-neutral; vertical suites such as mini-program engineering extend it with domain facts and platform rules.

Compare before adopting

Before replacing your existing handoff checklist or PR template, compare three things: quality signals (are issues and commits maintained?), freshness (is the README updated for current agent patterns), and workflow fit (does it integrate with how your agent stores session state and reports?). Alternatives include simple prompting rules and custom guardrails, but they usually lack the explicit evidence-status model and resumable recovery this skill formalizes.

Why it is worth tracking

Evidence-first engineering addresses a real pain point: agent output that reads confidently but cannot be verified. With 22 stars and a clear path to vertical suites, it is small but focused enough to be worth testing when you need auditability in agent-driven changes. For more context on how it compares with other automation skills, its listing at https://www.openagentskill.com/skills/nocodemrli-evidence-first-engineering is a useful reference when evaluating adoption.