system1-agents

[!NOTE]
Give your agents a System 1 decision model. Start from a prebuilt agent or build your own.

Describe the task. Claude Code or Codex runs a prebuilt agent or builds a new one, on Jev,
Laya or Cua-S1 Nano.
Browser use, computer use, robotics and games ship ready to run.

Up to 6× faster and 25× cheaper than a chat model, at the same score.

Quickstart · From Claude Code or Codex · Benchmarks · Docs

† The first Allrecipes task of the WebVoyager task set
(He et al., 2024, Apache-2.0, attribution in NOTICE): a vegetarian
lasagna with over 100 reviews, 4.5 stars or more, for 6. The chat model of that row is Claude Fable 5.1 through
OpenRouter; both models pay it for the typed search text and the answer. * Estimated; the chat-model run recorded no
cost. Each replay below is the episode behind its row, Jev on the left and the chat model on the right, both on the
wall clock. The other Allrecipes runs, longer games and the Google Flights driver comparison:
docs/benchmarks.md.

Browser use, WebVoyager's Allrecipes task 0 on the live site, replay at 8× speed	Computer use, the Windows Calculator, replay at 4× speed
Ticket router, 30 tickets to five queues, replay at 8× speed	ALFWorld, an embodied household task, replay at 2× speed
2048, 20 moves, replay at 8× speed	Millionaire, a 15-question quiz ladder, replay at 4× speed
Blackjack, one hand, replay at 2× speed

Choose your path

From Claude Code or Codex

npx skills add ThinkFlowLab/system1-agents                                                              # the skill: Claude Code, Codex, Cursor
claude plugin marketplace add ThinkFlowLab/system1-agents && claude plugin install s1a@system1-agents   # plus the browser subagent and the MCP server

The skill tells the host when to hand a task to a System 1 agent and which command to run. The ticket router from
the table above, from Claude Code:

Route this ticket to logistics, payment, returns, account or human: "I was charged twice for order 4411 and I
want the second charge refunded."

The host runs one s1a decide over the five queues and reports the queue with its probability, in about 400 ms.
Page tasks, games and one-off selections go the same way; arithmetic, deduction and free text stay with the chat
model. The walkthrough, the plugin's keys and what to delegate: docs/skills.md.

From the command line

git clone https://github.com/ThinkFlowLab/system1-agents && cd system1-agents
uv sync && cp .env.example .env     # the first sync resolves the openjiuwen pin and takes a few minutes

Put a Jev key in .env (TYPESAFE_API_KEY from the TypeSafe console, or
OPENROUTER_API_KEY), then ask for one decision and run one agent with each model:

uv run s1a decide --state '{"player_total": 18, "dealer_upcard": 9}' \
  --option hit="take a card" --option stand="keep the hand" --rules "stand on 17 or more"
uv sync --extra blackjack
uv run s1a run blackjack --model jev --rethink off --episodes 20
uv run s1a run blackjack --model llm --rethink off --episodes 20     # the chat model in the same agent

decide prints one JSON object with choice, a probability per option, confidence and ms; run writes a job
folder with the score. Without a key, --model cua answers in process after uv sync --extra cua.

As an MCP server

codex mcp add s1a -- uv run --project /path/to/system1-agents s1a-mcp

It serves three tools, list_agents, run_agent and decide, to any host that speaks MCP.

Build your own

A System 1 agent is one module under s1a/agents/ that ends in a frozen SPEC; Blackjack is 111 lines. The builder
skill runs in Claude Code from this checkout, probes the task with 8 to 12 hand-written decisions before it writes
code, and stops when the task needs deduction or arithmetic:

Build a System 1 agent for .

The gates and the templates: docs/skills.md.

What ships

• allrecipes: browser use, WebVoyager's first Allrecipes task, headed on the live site.
• flights: browser use, a Google Flights search over @playwright/mcp.
• desktop: computer use, any Windows or macOS app window through Cua Driver.
• ticket_router: 30 labelled support tickets to five queues.
• alfworld: household tasks in text, with the AI2-THOR scene in the replays.
• game2048, millionaire, blackjack: games with a score per episode.
• injection_guard: a rail that answers one question at a hook of a running agent and fails closed.

Every agent runs on jev, laya or cua, and on the chat model for the comparison. Flags, run commands and
extras: docs/agents.md.

How it works

Each agent is a stock openJiuwen agent with a System 1 decision
model as its model. On a decision turn the model gets the state and the options and answers with one of them;
planning, typed values and the final answer stay with the chat model in the same agent. Any decision model with that
interface fits: docs/architecture.md, docs/decision-models.md.

Docs

• docs/benchmarks.md: the six runs above, the Google Flights driver comparison, a longer game, the guard rail.
• docs/skills.md: the caller skill, the builder skill, what to delegate.
• docs/agents.md: every agent with its flags, run command and extra.
• docs/architecture.md and docs/decision-models.md: the fronts, the model slot, the model interface, adding a backend.
• docs/browser-front.md: the browser policy, decision by decision.
• docs/why.md: the problem, the philosophy, the precedents.
• docs/roadmap.md and CHANGELOG.md.

Contributing and license

CONTRIBUTING.md has the dev install, the checks and the hooks. Apache-2.0.