Evaluation RuleDecision layer

When to Adopt Coral Bricks: Your Client Runs Their Own Agent Loop

Should an agency adopt Coral Bricks as the inference layer for a client's coding or research agent, or is the client better served by a frontier model API? Adopt Coral Bricks only when the client owns an agent loop you can repoint at an OpenAI-compatible endpoint, and price the engagement on setup plus usage rather than a flat retainer.

By InnovaAI ResearchPublished Updated

Should an agency adopt Coral Bricks as the inference layer for a client's coding or research agent, or is the client better served by a frontier model API?

Adopt Coral Bricks only when the client owns an agent loop you can repoint at an OpenAI-compatible endpoint, and price the engagement on setup plus usage rather than a flat retainer.

Common Mistake

Treating Coral Bricks as a white-label product to resell to clients who have no agent of their own. The verdict states plainly that you cannot resell it as a standalone service, so agencies that quote a monthly retainer for 'AI' without an existing client agent loop end up owning an inference bill with nothing to attach it to.

Why This Works

Coral Bricks is an inference API for open models with pay-per-token pricing, for example $1.12 per 1M input tokens on GLM 5.3 and $1.68 per 1M cache write tokens, plus free cached reads that reward repeated context. The verdict is explicit that this is a developer-infrastructure play, not a client-facing SaaS platform, so resale works only if your clients are building or running their own agent systems. The published Starter Agent offer at $2,500 with 20h setup is the shape of a viable engagement: a fixed build fee on top of pass-through token usage.

Apply When
  • The client already operates a coding or research agent tool (OpenCode, Codex CLI, GitHub Copilot, Cursor, Cline, Continue, or aider) and can be repointed at an OpenAI-compatible endpoint
  • The agent loop repeatedly sends the same long context, so free cached reads materially cut the token bill
  • The client's workload fits served open models (GLM, Kimi, gpt-oss, DeepSeek) rather than requiring a proprietary frontier model
  • The engagement is a build or retainer where the agency owns the inference configuration, not a one-off prompt job
  • The client has a compliance reason to avoid vendor lock-in to a single proprietary LLM API