Awesome AI Agents › AI Assistants

Xiangyue-Zhang/auto-deep-researcher-24x7

⭐ 1292 Python repository created 2026-04-08

Deep Researcher Agent is a Python 3.10+ agent that runs deep learning experiments without a human in the loop, cycling through planning, execution, monitoring and reflection. It is organised as a leader-worker architecture with bounded, constant-size memory so token growth stays flat over long runs; leader history is reset and a no-progress fallback keeps the loop from stalling. Models can be driven through subscription-backed claude_cli and codex_cli provider modes or through any OpenAI-compatible endpoint, including one-word presets for DeepSeek, Qwen via DashScope, Kimi via Moonshot and GLM via Zhipu that auto-fill base URL and key environment variable while leaving both overridable for self-hosted proxies. Execution backends cover local runs, a remote SSH host where only the controller stays local, and Slurm clusters where training is submitted with sbatch over a single transient SSH call, sacct is the sole liveness authority and bounded checks guarantee the monitor loop terminates without reaping jobs that are still queued or running. A finished job reports its real terminal state, so failed, timed-out or cancelled runs are no longer recorded as completed. The autonomy layer adds an append-only experiment ledger in JSONL that records each cycle hypothesis, metrics and outcome and feeds back into planning; a data-driven stagnation signal derived from the metric trajectory; append-only DEAD_ENDS and INSIGHTS journals rotated rather than compacted; a zero-cost violation scanner with an advisory phase gate; and an optional cap on cycles per hour to protect budget. Agent tools include regex code search, a depth-limited repository tree, line-ranged file reads that stay inside the workspace, plus Semantic Scholar search, arXiv search and paper detail lookup with reference and citation snowballing. All configuration sections default to previous behaviour.

https://github.com/Xiangyue-Zhang/auto-deep-researcher-24x7

autonomous-agentresearch-automationllm-agentexperiment-trackingpython