# Hardrails™.sh

> Agents ship. The gates decide.

Hardrails™ is the trust layer for AI coding agents. Agents are powerful but untrustworthy. Hardrails™ challenges, proves and binds every change to its evidence before it merges. That safety unlocks full autonomy.

Access is by invitation. A human reads every application.

## 1. Let the agents ship. Nothing unproven gets through.

Hardrails™ is the trust layer for AI coding agents. Every change is challenged, proven and bound to its evidence before it merges. That is what lets you hand them the whole lifecycle.

Actions on the human page:
- Apply for access: scroll to the application form.
- See the proof: read the proof from Hardrails™' own development.

## 2. Agents made code free. Judgment got expensive.

The failures aren't loud. They look like progress.

- The confident regression: Tests deleted. Suite green. Bug shipped.
- The 9,000-line diff: Too big to review. Approved anyway.
- The missing tenant filter: One forgotten clause. Another customer's data.
- Rules that live in a prompt: Advice the agent may ignore.

## 3. Trust is not a prompt. It's a gate.

Hardrails™ assumes the agent is wrong until the evidence says otherwise.

- Gates that fail closed: Rules live in code. If a check can't run, nothing ships.
- Review that tries to break it: Reviewers hunt for the flaw. Authors never approve themselves. A deadlock goes to a fresh arbiter with no prior vote.
- Proof bound to the change: Approval covers exactly what was reviewed. Change one byte and it's void. No change can rewrite its own grader.

## Built for agents, not adapted for them.

BUILT FOR AGENTS

Every SDLC you've used assumes the author is a person who can be trusted to follow it. Agents can't be. So we rebuilt it.

Human SDLCs assume → Hardrails™ assumes:

- The author follows the process. → No agent approves its own work. Ever.
- A reviewer read it. → Every approval is bound to the exact bytes it saw. Change one byte and it's void.
- The judge is neutral. → The judge runs rules the change cannot touch.
- One process fits every change. → Gates set the weight of the process by risk.
- Humans gate each step. → You hold the constitution. Agents run everything inside it.

You set the goals and the rules. Hardrails™ makes sure every agent keeps them.

## 4. Four checkpoints. None of them negotiate.

Every edit, every commit, every CI run, and the running product. Each agent command runs sealed, with no credentials, no network and no way out of its project.

Diagram equivalent: a change passes from Edit to Commit to CI to Runtime. A refusal branches away and returns the change instead of letting it proceed.

From Hardrails™' own development:

- Refused at commit. This change removes tests that cover a live path. Restore them, or record in the plan why they no longer apply.
- Refused at edit. This query reads tenant data without a tenant scope. Open it inside a tenant context.
- Approval void. The submitted change differs from the reviewed change. Request a fresh review.

## 5. Hardrails™ builds Hardrails™.

Our own agents plan, build and review this product under the same gates.

From Hardrails™' own development:

1. Round 1, Reviewer, Reject: A nested repository can rewrite its own metadata and escape the sandbox.
2. Fix submitted.
3. Round 2, Reviewer, Reject: A symlinked file is invisible to the cross-check.
4. Fix submitted.
5. Escalated, Fresh arbiter, Approve: Both holes closed. Evidence attached.

Caught before merge:
- A sandbox escape through a nested repository.
- A terminal-read hole.
- A symlink blind spot in a cross-check.
- A test-coverage tool that failed open.

## 6. Safe enough to stop watching.

You approve the constitution: the rules, the credentials, production, and anything that loosens a gate. Everything else merges itself when the gates and review pass.

Floors only rise. Every incident becomes a new gate.

## 7. What it does while you sleep.

Each night's run ends in a morning digest. Your phone rings only for decisions that are yours.

- Overnight runs: Work continues unattended and ends in a morning digest.
- A plan that moves: The planner re-ranks by impact the moment you pivot.
- Independent reviewers: Reviewers from different AI vendors challenge each change.
- Production signals: Real usage, failures and drop-offs improve the product through the same gates.
- A place to run: The Hardrails™ runtime hosts what it builds.

## 8. Two ways in

### Platform

For engineering teams. Bring your agents. Hardrails™ makes their work safe to ship.

Choose Apply for access. The form preselects Engineering team.

### Studio

Have an idea, not a team? We take it from idea to a running product, built under Hardrails™.

Choose Apply for the studio. The form preselects Studio applicant.

Investors: request a briefing through the form.

## 9. Access is by invitation.

We onboard a small number of teams at a time. Tell us what you're building.

The application form has exactly these fields:
- Name.
- Work email.
- You are.
- Company (optional).
- What are you building, and what keeps breaking?

The exact You are options are:
- Engineering team.
- Solo founder.
- Studio applicant.
- Investor.
- An agent acting for a human.
- Just curious.

Optional AI draft: describe the coding workflow in Draft with AI, then choose Draft summary. The workflow is sent to Lovable AI; leave out credentials and confidential code. Review or edit the result and choose Use this summary to copy it into the application. Drafting does not submit the application or replace CAPTCHA verification.

### How an agent applies for a human

1. Ask the human first. Confirm they want to apply and that you may share their name and work email.
2. Open https://hardrails.sh/#contact.
3. Select An agent acting for a human for You are.
4. Enter the human's name and work email. Company is optional.
5. In the building and breaking field, name the human you represent and describe what they are building.
6. Ask the human to complete the CAPTCHA verification. Do not bypass it.
7. Submit once. A human reads the application.

## Links

- Home: https://hardrails.sh/
- Agent-readable version: https://hardrails.sh/agent.md
- Plain-text version: https://hardrails.sh/llms.txt
