Honest infrastructure for AI agents that outlive the session

What it does

Deuz SDK is a TypeScript framework for running AI agents in production with zero runtime dependencies. It wraps model calls in durable execution loops that checkpoint state to your own database, manage long-term memory with reconciliation and expiration, compact overflowing context windows automatically, and enforce guardrails and human approvals without phoning home to a vendor. It runs on Node, Bun, Deno, and edge runtimes.

The interesting bit

The entire runtime is built around a single design rule: normalize every provider’s bytes into a canonical delta stream first. That means retry, failover, resume, token budgets, and sub-agents all speak one language, and no raw provider SSE ever leaks to your caller. It also means the code is aggressively seam-injected—clock, randomness, fetch, and logging are all passed in—so tests stay deterministic and the same bundle runs from serverless to edge without Node leaks.

Key highlights

• Zero runtime dependencies; the core is one package with nothing ambient.
• Long-term memory pipeline that extracts facts, reconciles them against existing knowledge, and recalls from vector stores, Postgres, or even an Obsidian vault.
• Automatic context compaction: prunes stale tool output and folds old turns into a running summary when the window fills, with force-compaction retries if a provider rejects a long request.
• Crash recovery via step checkpoints in your own database—resumeFromCheckpoint later with no workflow vendor lock-in.
• 28 chat providers normalized behind one streaming API, plus MCP tool servers with OAuth 2.0, human-in-the-loop approval tokens, and handoffs between agents.

Caveats

• Overflow recovery does not yet work on the Gemini native wire, and generateObject cannot coerce DeepSeek V4.
• Token counting is a calibrated heuristic unless you bring your own tokenizer; speech, transcription, and video are currently covered by mocked tests but not yet verified against live endpoints.
• The Redis pack lacks MULTI, MCP has no WebSocket transport, and the Part union is missing AudioPart.

Verdict

Worth a look if you want a small, hold-in-your-head runtime that keeps state in your database and refuses to phone home. If you need the largest ecosystem and integrations today, the README itself admits the Vercel AI SDK is the safer bet.