By ByteBot
Most TypeScript agent frameworks fail the same way: the agent runs fine until a deploy kills it mid-step, a network drop loses 40 minutes of work, or a crash reruns a tool call that already sent an email or charged a card. Deuz SDK 2.4 — released this week and trending on GitHub — ships the two features production agents have been missing: memory that survives restarts and background execution that resumes after a crash without replaying side effects.
What Deuz SDK Is
Deuz SDK is a TypeScript agent framework where @deuz-sdk/core ships with zero runtime dependencies. Clock, randomness, fetch, keys, and logging are all injected rather than bundled — which means the same code runs unchanged on Node, Bun, Deno, serverless, and edge. You get one streaming API covering 29 providers including Claude, GPT, Gemini, Grok, DeepSeek, Mistral, and Ollama. Swap models without touching your code structure.
Versions 2.1 through 2.3 built out native agents, durable swarms, and single-session reliability. Version 2.4 solves the harder problem: agents that run over hours or days, not minutes.
The Two Features That Actually Matter in 2.4
Goal Supervisor: Background Execution That Survives Crashes
The goal supervisor runs agent goals on a cron schedule — with proper timezone and daylight saving handling — and from verified webhooks. The important bit is what happens when the process dies: the supervisor resumes the goal from where it stopped, without repeating external effects and without double-charging usage.
If your agent was on step 8 of 10 and had already called Stripe, filed a Jira ticket, and sent a Slack message, a crash-resume does not redo those. It picks up at step 9. That is the “no duplicate Stripe charge” guarantee that most teams are currently building by hand — badly — or not building at all. A 2026 survey found that 42% of teams reported revenue loss from agent failures. The category of failure being solved here is real.
Versioned Memory: Facts With History and Redaction
Deuz 2.4 replaces “here is some memory” with versioned memory: every fact is stored alongside its history and source attribution. When you call forget(), it does not just delete the fact — it redacts the stored context that held it. That matters for GDPR, for compliance, and for any situation where an agent accumulated incorrect context that needs to be cleanly removed rather than shadowed.
A context compiler handles the practical side: it fits memory, documents, past chats, and loaded skills into a single token budget, maintaining a working memory for the current goal, constraints, and decisions. Transcript search finds past messages across sessions. Backends are your choice: vector stores, Postgres, or Obsidian vaults.
Quick Start
npm install @deuz-sdk/coreimport { runAgent } from '@deuz-sdk/core/agent';
import { createAnthropic } from '@deuz-sdk/core/anthropic';
const anthropic = createAnthropic({ apiKey: process.env.ANTHROPIC_API_KEY });
const result = await runAgent({
model: anthropic('claude-opus-4-8'),
prompt: 'Investigate this production incident and file a ticket.',
maxSteps: 10,
});
if (result.status === 'completed') console.log(result.output);The agent pattern is intentional: runAgent returns a typed result with an explicit status. No undifferentiated exceptions from a black-box runner. For step checkpoints, pass a database adapter and every completed step is persisted automatically.
Where It Sits vs. the Competition
Mastra is the closest competitor: TypeScript-native, durable execution, strong type safety, and human-in-the-loop steps. But Mastra carries runtime dependencies. Deuz wins on two things Mastra cannot currently match: the zero-dependency guarantee and true cross-runtime compatibility — the same bundle running on Node, Bun, Deno, and edge without polyfills or environment checks.
Vercel AI SDK is still the right choice for frontend-heavy streaming UIs with React. Its edge-native architecture processes streaming tokens 2-3x faster at time-to-first-token. It does not ship durable execution natively. LangChain.js has the most integrations — 750+ — and is primarily Python-first. It adds 15-40ms overhead per chain step. Use it when you need an obscure integration; do not use it as your default TypeScript agent primitive in 2026.
Also in 2.4
A local Docker sandbox for agent-written code is in preview — the first time Deuz ships first-party code execution. Connection continuity means dropped connections do not terminate agent responses. Parallel tool calls include self-healing errors and runaway guards (cost and token budgets). The full feature list also includes human-in-the-loop approvals with HMAC-signed expiring tokens, MCP client/server, and Agent-to-Agent (A2A) protocol support.
Who Should Use It
TypeScript teams deploying production agents — not prototypes — especially those running across multiple environments. If you are already on Vercel AI SDK and the work is a streaming chat UI, stay put. If you are building an agent that runs on a schedule, calls external services, waits on human approvals, and needs to keep working after a deploy: evaluate Deuz 2.4.
The zero-dependency constraint makes it the default choice for teams shipping to edge runtimes where dependency weight matters. For Bun-first shops in particular, the cross-runtime neutrality is genuinely useful — most frameworks still treat Bun as a second-class runtime.
Install with npm install @deuz-sdk/core. The full documentation and changelog are on GitHub.
