jagnani73P.03
[the record]STATUS: NTU SINGAPORE — AUG 2026

AI Agent SDK

OPEN SOURCE · 119★
the orchestration layer behind Covalent's Zero-Employee Enterprise — agents that read chains, call tools, and hand work to each other
fig. 1 — zee.run(): the planner breaks the goal into tasks, the router assigns them, the endgame concludeslive trace
goal inAGENT · plannerAGENT · routerTASK AGENT · 1TASK AGENT · 2TASK AGENT · 3AGENT · endgameanswer
01THE PROBLEMeveryone rebuilding the same glue

LLMs could finally reason about on-chain data — but every team was hand-rolling the same fragile glue to let them act on it.

ZEE — the Zero-Employee Enterprise — needed an orchestration layer where autonomous agents could interact with on-chain data, external tools, and each other, without each integration being bespoke. I built and shipped it as @covalenthq/ai-agent-sdk on npm — 119 stars, 56 forks, and 2,000+ downloads since its December 2024 launch.

02THE ARCHITECTUREthree default agents around yours · TypeScript

A ZeeWorkflow wraps the task agents you define with three defaults. The planner breaks the goal into dependency-ordered tasks; the router assigns each task to the most suitable agent — and answers their follow-up questions mid-run with full context; the endgame agent compiles everything into the final response. An action queue drives the whole run, capped at 50 iterations.

01

PLAN

the goal becomes dependency-ordered tasks

planner agent

02

ASSIGN

each task routed to the best-suited agent

router agent

03

ACT

your agents complete tasks; follow-ups loop back through the router

task agents · tools

04

CONCLUDE

the full context, compiled into one answer

endgame agent

STACK — TypeScript · Vercel AI SDK · OpenAI · Google · Anthropic · GoldRush API · npm
03WHAT I SHIPPEDcore SDK · launch → v0.3.0

I built the core SDK package and led the architectural milestones from launch through v0.3.0:

01

multimodal support

Gemini image analysis alongside text reasoning

02

Vercel AI SDK migration

one layer for streaming + tool calls — killed a cross-provider tool-calling bug

03

ZEE assignment refactor

the v0.3.0 agent-assignment model

04

cza CLI

scaffold a new agent project in one command

04IN THE WILDplates 01–02 · npm · github
import { Agent, ZeeWorkflow } from "@covalenthq/ai-agent-sdk";

const analyst = new Agent({
  name: "onchain-analyst",
  model: { provider: "OPEN_AI", name: "gpt-4o-mini" },
  description: "reads balances, transfers, and approvals",
  tools: { balances, transactions },
});

const zee = new ZeeWorkflow({
  description: "audit the wallet, then write the report",
  output: "a structured risk report",
  agents: { analyst, writer },
});

const result = await ZeeWorkflow.run(zee);