qwwiwi/agent-jobs

> Read in your language: English (this page) · Русская версия →
## One brief in. Finished work out.
Agent Jobs turns one Claude Code session into a team lead. You give the session a large task once. It writes briefs and hands the work to background claude -p jobs. Each job runs alone in its own git worktree, under hard limits and a guard, and comes back with a report. With the brief flow, the session checks the result against a checklist approved by the owner, and releasecheck passes the release only when every requirement is met. The owner answers the open questions and approves the checklist; everything else runs without them.
Cheaper. Your subscription limits go further overall, because the main session orchestrates and keeps its own context lean. On one production agent, priced usage from transcripts (real usage) went from ≈ 14 subagent tasks to ≈ 146 jobs a day at about twice the daily dollars, and the main session's context per call fell from 390–560k to ≈ 265k tokens (an earlier auto-compact change also helped): this suggests far less limit use per task, though jobs are smaller tasks and limit burn itself was not recorded. A Claude job's spend is capped per attempt, with a $50 ceiling across attempts, and a single small job is not cheaper per task than a subagent (benchmark).
Faster. The session gets its turn back at once instead of waiting on a worker. Jobs run in parallel slots and queue the rest, and they survive a restart of the session or the daemon.
Better. Every job is a full top-level Claude Code run with its own context and a fixed report contract. Large tasks are checked row by row against the owner's checklist, with evidence and red proofs, by a blind judge; releasecheck passes only a judge that is blind by enforcement and of a different model family from the coders.
Autonomous. Time limits on every job; on Claude jobs also turn and budget limits, a guard on every tool call and one automatic continuation after a turn-limit or timeout stop. Codex jobs run in the codex sandbox instead of the guard (read-only unless codexwriteallowed) and are neither continued nor resumed. Schedules and dead-man's switch checks run on their own. Jobs never push or publish: only the session does, after it has read the report and checked the diff.
It is a single Go binary with one SQLite file per agent. It runs on Linux (systemd user service) and macOS (launchd).
## Quick start
You need Go 1.27, git, the claude CLI and tmux (for the job-done line).
1. Install (Linux). One script builds the binary, installs the systemd user unit and the agent-jobs skill, initialises the agent and starts the daemon:
bash
git clone https://github.com/qwwiwi/agent-jobs && cd agent-jobs
scripts/install-linux.sh --instance my-agent
systemctl --user status agent-jobs@my-agent
On macOS, build with go build -o ~/.local/bin/agent-jobs ./cmd/agent-jobs, run agent-jobs init --agent my-agent and load the launchd template: see Install and configure.
2. Tell it where your session lives. In ~/.agent-jobs/my-agent/config.json, name the tmux target of your Claude Code session, so the job-done line reaches it:
json
{ "notifytmuxtarget": "my-agent-session" }