FAQ
General
Section titled “General”How do i start?
Section titled “How do i start?”Create a repository from apex-accelerator and follow the quickstart. APEX owns the product. This apex-docs repository owns the published site.
Is generated code ready for production?
Section titled “Is generated code ready for production?”Not by generation alone. Review requirements, architecture, policy, the implementation plan, generated code, and the actual deployment preview. Complete the required checks for your environment. Record application health separately from resource provisioning. The historical case study shows why that distinction matters.
Which Copilot plan and model do i need?
Section titled “Which Copilot plan and model do i need?”Use a Copilot plan and organization configuration that provide the required custom
agent and model access. Model declarations are in .github/agents/*.agent.md.
They do not grant access to a model. Check your actual entitlement and
GitHub’s plan comparison.
This site does not require Business or Enterprise by assertion.
Do i need Azure access?
Section titled “Do i need Azure access?”You can explore prompts and draft requirements without a subscription. Public retail pricing does not require scoped cost-management access. Effective policy discovery, target-specific validation, deployment, and live diagnostics require appropriate Azure access. Offline drafts are not fully verified deployment inputs.
IaC tracks
Section titled “IaC tracks”Should i choose Bicep or Terraform?
Section titled “Should i choose Bicep or Terraform?”Bicep is Azure’s resource-definition language and uses ARM-managed deployment state. Terraform uses providers and a state backend. Choose according to your team’s operating model and governance requirements.
Requirements records iac_tool as Bicep or Terraform. Both tracks use
05-IaC Planner. CodeGen and Deploy have track-specific main agents, which the
owner selects. See the workflow.
Can i change tracks later?
Section titled “Can i change tracks later?”Treat it as a requirements change. Record the decision through the owning agent, review affected architecture and module choices, then regenerate the plan, contracts, code, and handoff as needed. Do not change one field and assume every prior approval remains valid. Existing Terraform state or deployed resources need an explicit migration plan.
Does APEX work offline?
Section titled “Does APEX work offline?”Copilot requires network access. Tool calls also depend on their service endpoints. You can read local files offline, but missing pricing, policy, or validation evidence must remain missing. Do not substitute remembered prices or call unavailable checks successful.
What is the difference between an agent and a skill?
Section titled “What is the difference between an agent and a skill?”A main agent owns a workflow step and its handoffs. A skill contains reusable instructions and references. Invocation flags control how a skill may be selected. Neither a skill nor source metadata grants additional tools or permissions.
See agents and skills and instructions.
How do i resume?
Section titled “How do i resume?”Select 01-Orchestrator, name the project, and ask it to inspect current state
and artifacts. It recommends the next handoff; you select the main agent.
Use session state debugging for corruption,
conflicts, stale indexes, or authorized recovery.
Is there a hands-on exercise?
Section titled “Is there a hands-on exercise?”The MicroHack is a separate learning resource. Check its environment requirements and revision before following deployment steps. The April 2026 case study on this site is a historical record, not a current copy-and-run tutorial.
Customization & multi-project
Section titled “Customization & multi-project”How do i customize an accelerator repository?
Section titled “How do i customize an accelerator repository?”A repository created from the template belongs to you. Inspect its current sync workflow before editing files you want to retain.
At the reviewed template revision, scheduled upstream sync proposes a pull request. Manual runs default to dry-run. The sync mirrors upstream content except for declared exclusions, shared-file exceptions, and seed rules. It does not merge local edits into upstream-owned files automatically.
Both generated Bicep and Terraform project directories are excluded, but their
shared AGENTS.md files are explicit sync exceptions. Workflow files remain
separate and use npm run sync:workflows. Root AGENTS.md, agents, skills,
and instructions are not generally protected customization locations.
Read the actual sync configuration and review each proposed diff. Do not rely on an old four-path exclusion list.
Can one repository contain multiple projects?
Section titled “Can one repository contain multiple projects?”Yes. Project artifacts live under agent-output/{project}/, and generated IaC
lives under the matching track’s project directory. These projects share agents,
skills, and container configuration. Separate repositories provide separate
permissions, lifecycle, and tooling configuration when your team needs them.
Troubleshooting
Section titled “Troubleshooting”Why does orchestrator not delegate to the next main agent?
Section titled “Why does orchestrator not delegate to the next main agent?”Check the selected agent’s permissions and the actual native task result.
The body-migration candidate permits 01-Orchestrator to dispatch approved main
workers; direct main-agent selection remains supported. The owner-approved
shared-review candidate permits automatic producers and authorized Orchestrator
calls to use challenger-review-subagent. It does not require manual
10-Challenger selection. Missing capability or invalid evidence stops progress,
without an automatic fallback. Do not add wildcard delegation permissions.
See the candidate contract and runtime limits.
What if an agent produces incorrect output?
Section titled “What if an agent produces incorrect output?”Provide the exact error, affected artifact, and expected behavior to the owning step. A deployment agent must return code defects to CodeGen. Preserve required reviews after changes. Starting a new chat does not make stale approvals current.
What if MCP is unavailable?
Section titled “What if MCP is unavailable?”Identify the failed server and required evidence. Work that does not depend on it may continue, but affected checks remain blocked or unperformed. Cost estimates must not substitute invented prices. Azure CLI authentication for governance is separate from pricing-tool availability.
Where do i report problems?
Section titled “Where do i report problems?”Use APEX issues for product defects and apex-docs issues for site problems. Include the revision, failing command or agent, and redacted evidence. See troubleshooting for diagnosis.