Skip to content

agents-as-code.md

Latest commit

 

History

History
31 lines (19 loc) · 2.46 KB

agents-as-code.md

File metadata and controls

31 lines (19 loc) · 2.46 KB
Raw

Agents as code

The problem

Cloud AI agents are increasingly built by clicking through provider consoles: create an environment here, upload a skill there, paste an API key, wire up an MCP server. The resulting state lives only in the console — it cannot be reviewed in a PR, rolled back, reproduced on another account, or moved to a different vendor. As agents move from "call a model API in your own code" to "assemble an agent inside a vendor console", vendor lock-in re-forms one layer up — at the managed harness.

The bet

OpenAgentPack treats an agent as infrastructure as code, the way Terraform treats cloud resources. A single agents.yaml is the desired state and the single source of truth. A Terraform-style workflow reconciles the real provider to match it:

validate → plan → apply → destroy

Because the declaration is a file, it gets everything a file gets: code review, version history, branching, rollback, and reproducibility across accounts and providers.

Two layers worth distinguishing

  • The agent harness is the provider-managed layer that wraps a model into an agent: knowledge base, skills, MCP wiring, prompt/instructions, vault, deployment, multi-agent orchestration. These are the customer's portable assets.
  • The agent infra is the interchangeable execution substrate beneath the harness — the specific provider (Bailian, Qoder, Claude, Volcengine Ark) that runs the agent.

OpenAgentPack's portability claim is that the same harness declaration can target different agent infra. Portability means the core declaration is portable and the per-provider capability contract is explicit — unsupported facets degrade gracefully (for example, an emulated Deployment on Bailian/Volcengine Ark) — not that every feature is identical on every provider.

What this enables

  • Preview before mutation — plan shows every create / update / delete before any API call mutates remote.
  • Incremental runs — content-hash diffing skips unchanged resources; only what actually changed is touched.
  • Dependency-aware ordering — Environment → Skill → Agent are created in topological order; a failed dependency skips its dependents instead of leaving half-built state.
  • Drift recovery — because the YAML is the source of truth, apply reconciles remote drift back to the declaration.

Continue with Resources and State and drift for the mental model behind these behaviors.