@atbash/autogen

Atbash safety judge for AutoGen-style multi-agent orchestration loops.

This package is intentionally small. It gives you one focused helper to ask Atbash for a verdict at the point where your app decides whether to proceed with an action. It does not own your orchestration model.

Installation

npm install @atbash/autogen

When To Use It

Use this package when:

• you already control your own orchestration steps
• you want one explicit Atbash check before a side effect
• you do not need a heavier plugin lifecycle

Good fits:

• AutoGen-style multi-agent loops
• custom planners
• supervisor-worker systems
• approval chains where your app already owns the review UI

Quick Start

import { createAtbashClient, loadAgent } from "@atbash/sdk";
import { judgeForAutoGen } from "@atbash/autogen";

const agent = loadAgent(process.env.ATBASH_AGENT_PRIVKEY);
const client = createAtbashClient({ keyPair: { privKey: agent.privkey, pubKey: agent.pubkey } });

const result = await judgeForAutoGen(
  {
    action: "Bank transfer $25 to a new external vendor account",
    context: "AutoGen agent checking transfer before execution",
    toolName: "send_bank_transfer",
    toolArgs: { amount: 25, recipient: "new vendor" },
  },
  client,
);

if (result.allow) {
  // proceed
} else {
  // stop — surface result.reason to the operator
}

API

judgeForAutoGen(input, client)

| Parameter | Type | Description | |---|---|---| | input | AutoGenJudgeInput | The action to evaluate | | client | AtbashClient | SDK client created with createAtbashClient() |

Returns Promise<Decision>.

AutoGenJudgeInput

| Field | Type | Required | Description | |---|---|---|---| | action | string | Yes | Human-readable description of the action | | context | string | Yes | Why the agent is taking this action | | toolName | string | No | Name of the tool being called (defaults to "autogen_action") | | toolArgs | unknown | No | Structured payload the judge evaluates (defaults to { action }) |

Decision

| Field | Type | Description | |---|---|---| | allow | boolean | Whether to proceed | | verdict | "ALLOW" \| "HOLD" \| "BLOCK" \| "ERROR" | Judge verdict | | reason | string? | Policy reason (present on HOLD/BLOCK) | | toolCallId | string? | ID to pass back on HOLD resolution |

Verdict Handling

| Verdict | Meaning | Action | |---|---|---| | ALLOW | Safe to proceed | Continue orchestration | | HOLD | Needs human review | Stop and hand off; keep toolCallId | | BLOCK | Policy violation | Stop and surface reason | | ERROR | Judge unreachable | Fail closed by default |

Creating the Client

Create the AtbashClient once at startup, then pass it to every judgeForAutoGen call.

import { createAtbashClient, loadAgent } from "@atbash/sdk";

const agent = loadAgent(process.env.ATBASH_AGENT_PRIVKEY);
const client = createAtbashClient({ keyPair: { privKey: agent.privkey, pubKey: agent.pubkey } });

To use a custom endpoint:

const client = createAtbashClient({
  keyPair: { privKey: agent.privkey, pubKey: agent.pubkey },
  judge: { endpoint: process.env.ATBASH_ENDPOINT },
});

What This Package Does Not Do

• It does not wrap your framework for you.
• It does not create a review queue.
• It does not log or execute the real action automatically.

That is intentional. The host loop stays in control.

Example

A runnable example is in examples/autogen-runtime-agent/.

cd examples/autogen-runtime-agent
ATBASH_AGENT_PRIVKEY=your_key_here node run.mjs