Why reach for this instead of Claude Code or Codex alone?

Both are strong at single turns like edits and answers; this adds the layer above with cross-session memory, persisted turns, poll and wait primitives, and per-turn model routing for long-lived jobs.

How does resume survive a dead process or a new container?

Only TurnState must survive between turns: it is persisted on disk, then a fresh sandbox, serverless invocation, or different machine resumes it, including week-long waits for external replies.

What does the model router actually decide per turn?

You pick a tier, not a model: a cheap classifier picks the concrete model and effort per turn, keeps rechecking as work shifts, weighs prompt-cache loss, and can escalate consequential calls to an advisor.

Do I need Anthropic and OpenAI keys, or does one duet login cover it?

One login covers it. npm install -g @duetso/agent, then duet login writes DUET_API_KEY to ~/.duet/.env and unlocks Claude, GPT, Gemini, image and video models, and Firecrawl-backed web skills through the Duet AI Gateway, so there is no separate provider billing. Bring-your-own-keys setup is documented as the alternative if you would rather hold your own credentials.

What are relays in duet-agent?

Relays are agent-routed state machines with five state kinds: agent, script, poll, timer, and terminal, so a job can sit in wait_for_reply for a week without a process alive.

How does memory work across sessions?

The agent records observations from its own transcripts into PGlite with hybrid retrieval, pgvector plus tsvector fused via RRF, so it can recall them across sessions, including text observations made from screenshots and UI captures.

An optional escalation valve that tiers can enable; it injects a no-parameter ask_advisor tool that ships a curated slice of the transcript to a stronger model, and consulting is mandatory for consequential architecture calls, conflicting constraints, and important unknowns.

CommunityJev Agents & MCP

Duet Agent

Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.

See highlights

Open github.com

Stars: 48
Language: TypeScript
License: APACHE-2.0
Published by: Community
Author: dzhng
Last verified: Oct 2026

dzhng/duet-agent

User repository on GitHub

View on GitHub

An opinionated full-stack agent harness with native memories, long running tasks, and multi-agent relay

Forks: 4
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

The agent harness for jobs that outlive the chat. How do you keep an agent working until the job is done? Most harnesses don't have an answer — the chat ends, the process dies, the context goes with it. duet-agent has three. A session can pause for minutes or months and pick up in a fresh sandbox without losing the thread.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Routing Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Persists TurnState between turns so a job can pause for days and resume in a fresh sandbox, serverless invocation, or different machine.
Runs relays as agent-routed state machines with agent, script, poll, timer, and terminal states, including long waits between turns.
Records transcript observations into PGlite with hybrid vector plus full-text retrieval, including text observations from visual inputs.
Uses a classifier to pick model and effort per turn with cache-aware routing, installed via npm with one login key for frontier models.

Watch out

Built for jobs that span sessions and processes; work that fits one chat likely does not need it. Bring-your-own-keys setup is documented as an alternative. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

skilled-agent-harness_spec-driven-loops

Assistant framework that stores specs, memory, and specialist agents as plain markdown files in your repo for Claude Code, Codex, Opencode, and similar runtimes.