Phase 0 — foundation & contracts
codenzia/agents currently ships the core contracts and the
experiment harness. The concrete planner loop is being harvested from production agent code. This page previews the
target API — register for updates on the website to hear when it ships.
Codenzia Agents
The PHP-native LLM agent runtime behind the Agentic SaaS in PHP initiative —
declarative agents, an auto-discovered tool registry, and a bounded planner loop. Built on
codenzia/laravel-ai-assistant.
Define an agent as instructions + an allowlisted tool set + a model policy. No bespoke loop wiring per use case.
Tools are capabilities with JSON-schema params and an execute() body, discovered per module and gated by your tenancy / RBAC rules.
The AgentRunner drives a think → tool → observe loop with a hard step cap, so runaway tool-calling cannot happen.
Every run is logged as an AgentRun (steps, tokens, cost) — the unit you meter and bill against.
Built-in A/B exposure + conversion logging for agent surfaces, with an agents:experiment report for assisted-vs-unassisted lift.
Haiku for classify/route, Sonnet for the default loop, Opus for hard reasoning — OpenAI/Gemini stay available for portability.
Model policy
Haiku 4.5
classify / route / extract — lead scoring, intent.
Sonnet 4.6
default agent loop — shopper, support, content.
Opus 4.8
hard reasoning — pricing, strategy.
OpenAI / Gemini drivers stay available for portability.
Define a tool
A tool is a capability with JSON-schema params and an execute() body, discovered per module and gated by your RBAC.
Declare an agent
Instructions + an allowlisted tool set + a model policy. The runner handles the think → tool → observe loop with a hard step cap.
A/B experiment harness
The one piece shipping today: exposure + conversion logging for agent surfaces, with a report command for assisted-vs-unassisted lift.