Skip to main content
Hyde Workshop Human-governed AI

AI Forge // Controlled implementation

AI Forge: AI deployment controls before production.

AI Forge turns an AI idea, an automated workflow, or a software-agent task into a governed implementation plan. It sets out mission, evidence, permissions, observability, human approval, and rollback before anything consequential ships.

Build a deployment contract Review the six controls

Also see the AI Lab reviews and the governance workshop.

AI Astra, the accountability agent, representing human-approved AI deployment in the Hyde Workshop AI Forge

Three lenses on every deployment.

Each proposal is tested for what could fail and for what is worth keeping. A.M.A.N.D.A. then recommends a route.

Cain // Failure pressure

  • Unsupported claims and stale evidence
  • Broad credentials and hidden actions
  • Private data crossing boundaries
  • No independent stop or rollback

Abel // Practical value

  • Useful, measurable output
  • Repeatable work with artifacts
  • Less friction, authority retained
  • A clear maintenance cost

A.M.A.N.D.A. // Governed route

  • Adopt low-risk assistance
  • Pilot uncertain capability within bounds
  • Restrict permissions until proven
  • Reject if ownership is unclear

Six AI deployment controls must be in place.

Lead: AI Amy

Mission control

Define the measurable outcome, target user, operational scope, acceptance criteria, exclusions, budget, and the specific actions that must stay outside automated authority.

Required artifact: mission brief, acceptance criteria, prohibited actions, and a named human owner.

Lead: AI Maya

Evidence base

Verify the product capability, current version, official documentation, data sources, conflicting claims, known limitations, and unresolved evidence gaps before integration.

Required artifact: source map, version and date record, limitations ledger, and unresolved questions.

Lead: AI Nancy and AI Dawn

Strict permissions

Grant the system only the minimum read, write, send, spend, delete, account-change, filesystem, or deployment permissions the bounded mission needs.

Required artifact: permission matrix, credential boundary, and independent stop conditions.

Lead: AI Aria and AI Nancy

Observability

Keep enough logs, tool-call records, output traces, exceptions, handoffs, and acceptance tests to explain what happened and exactly where a failure started.

Required artifact: execution record, checkpoints, error pathing, and test evidence.

Lead: AI Astra

Human approval gate

Publishing, deployment, spending, deletion, customer communication, and security actions must stop at an explicit human decision. There is no autonomous deployment.

Required artifact: an APPROVE, REVISE, or REJECT receipt from an authorized human.

Lead: AI Dawn and AI Nancy

Rollback protocol

Define how to stop the system, restore the prior state, revoke credentials, recover data, revert code, and confirm that the rollback actually worked.

Required artifact: rollback tiers, recovery steps, a validation checkpoint, and a monitoring window.

Deployment contract builder.

Choose your workflow and consequence level. The six controls update as you change them. Everything runs in your browser.

Do not enter passwords, API keys, or customer data.

Your deployment contract

MCP server / agent integration · Medium consequence

  1. MissionDefine the agent's MCP server dependencies, prohibited API endpoints, budget limits, and success criteria. Require a checkpoint before any write.
  2. EvidenceVerify protocol specifications, identity provider, data handling terms, and known failure vectors.
  3. PermissionsRestrict tool scope; separate read capabilities from write, send, and spend authority. Keep credentials server-side. Require a checkpoint before any write.
  4. ObservabilityLog prompt context, MCP tool responses, handoff payloads, token usage, and blocked policy attempts.
  5. Human approvalRequire explicit human authorization before consequential writes, spending, or cross-system data transfers. Require a checkpoint before any write.
  6. RollbackProvide a kill switch, a credential revocation procedure, and a state recovery plan.

Questions before you deploy.

Is the deployment contract legally binding?

No. It is a planning checklist that makes the six controls explicit before anything ships. Have a qualified professional review any real agreement.

Does the builder send what I type anywhere?

No. It runs in your browser, makes no network request, and stores nothing. Still, do not enter passwords, API keys, or customer data.

Who decides whether to deploy?

A named human owner. AI Forge prepares the contract. Approval, publishing, spending, deletion, and security changes stay with that person.

How is this different from the Workflow Diagnostic?

The Workflow Diagnostic estimates how risky and how promising a workflow looks. AI Forge turns a workflow you have chosen into the six controls that should exist before it runs.

Next: see how the method works on the A.M.A.N.D.A. Analysis page, try the Workshop simulation, or send a project brief.