KwaiAgents

A lightweight information-seeking agent system runnable with GPT-3.5 or self-hosted KAgent models.

Stars
★ 1.2k
Last updated
2y ago
License
NOASSERTION
Primary language
Python

At a glance

Works with
Portable with changesOpenAI API
You'll need
Python 3.10+MinicondaOpenAI API key for GPT-3.5 usageShell / CLINetwork accessLocal filesystem
Typical use
A team that wants to handle English or Chinese information-seeking queries from a command line using GPT-3.5-turbo.
Main limitation
KAgentSys-Lite explicitly has a more limited tool set and no memory mechanism, so it is not equivalent to the full KAgentSys described in the paper.

What does this agent do, and when should you use it?

KwaiAgents is a collection of agent-related work released by Kuaishou Technology's KwaiKEG, centered on the runnable KAgentSys-Lite system. Its kagentsys CLI accepts a query and can use GPT-3.5-turbo or a deployed local KAgentLM to produce a response. The project also publishes KAgentLMs trained through Meta-agent tuning, the 200k-plus KAgentInstruct dataset, and KAgentBench with more than 3,000 human-edited evaluation examples. KAgentBench evaluates planning, tool use, reflection, concluding, and profiling, then reports dimension scores and an overall score. KAgentSys-Lite is not the paper's full KAgentSys: it has fewer tools, no memory mechanism, somewhat reduced performance, and a codebase evolved from projects including BabyAGI and Auto-GPT.

A user submits a question with kagentsys --query="..." --llm_name="gpt-3.5-turbo" --lang="en", and the system can call GPT-3.5-turbo using OPENAI_API_KEY. For local inference, KAgentSys-Lite can be pointed at an OpenAI-format service started with FastChat and vLLM by using --use_local_llm, --local_llm_host, and --local_llm_port; the Qwen GGUF conversion can also be served on CPU with llama_cpp.server. Weather requests can use WEATHER_API_KEY, while the browse_website tool requires chromedriver on the server. For evaluation, benchmark/infer_qwen.py writes JSONL inference results and benchmark/benchmark_eval.py reads them to report plan, tooluse, reflextion, conclusion, profile, and overall scores.

  1. A team that wants to handle English or Chinese information-seeking queries from a command line using GPT-3.5-turbo.
  2. A user who has deployed Qwen-7B-MAT, Qwen-14B-MAT, or Baichuan2-13B-MAT with vLLM and FastChat and wants to connect it to KAgentSys-Lite.
  3. A developer who needs to serve the Qwen-7B-MAT GGUF model through an OpenAI-compatible endpoint on a CPU-only machine.
  4. A researcher comparing models on planning, tool use, reflection, concluding, and profiling with quantitative KAgentBench results.
  5. An application developer handling weather-related agent queries who can configure WeatherAPI credentials.

How do you install or deploy this agent?

Install Miniconda, then run:

conda create -n kagent python=3.10
conda activate kagent
pip install -r requirements.txt
git clone [email protected]:KwaiKEG/KwaiAgents.git
cd KwaiAgents
python setup.py develop

For GPT-3.5-turbo, set:

export OPENAI_API_KEY=sk-xxxxx

For weather queries, also set:

export WEATHER_API_KEY=xxxxxx

How do you use this agent?

First GPT-3.5-turbo invocation:

kagentsys --query="Who is Andy Lau's wife?" --llm_name="gpt-3.5-turbo" --lang="en"

For a local model, start the FastChat controller, vLLM worker, and OpenAI API server:

python -m fastchat.serve.controller
python -m fastchat.serve.vllm_worker --model-path $model_path --trust-remote-code
python -m fastchat.serve.openai_api_server --host localhost --port 8888

Then run:

kagentsys --query="Who is Andy Lau's wife?" --llm_name="kagentlms_qwen_7b_mat" --use_local_llm --local_llm_host="localhost" --local_llm_port=8888 --lang="en"

What are this agent's strengths and limitations?

Pros
  • It combines a runnable KAgentSys-Lite implementation with KAgentLMs, KAgentInstruct, and KAgentBench, covering operation, models, training data, and evaluation data.
  • It supports both GPT-3.5-turbo through OPENAI_API_KEY and a connection to self-hosted local model services.
  • Its documented local serving options cover vLLM/FastChat on GPU and llama-cpp-python with GGUF on CPU.
  • KAgentBench reports planning, tool use, reflection, concluding, profiling, and overall scores for dimension-level model comparison.
Limitations
  • KAgentSys-Lite explicitly has a more limited tool set and no memory mechanism, so it is not equivalent to the full KAgentSys described in the paper.
  • Local-model use requires deploying FastChat and vLLM or installing llama-cpp-python; the GPU path may also require handling Bfloat16 compatibility.
  • Website browsing requires chromedriver on the server, and repeated search failures may require an http_proxy because duckduckgo_search may be unreachable.
  • The repository license status is NOASSERTION, so adopters need to establish the applicable licensing terms.

How does this agent compare with similar options?

KAgentSys-Lite is positioned as a lighter version of the full KAgentSys: it has a smaller tool set, no memory mechanism, somewhat reduced performance, and a different codebase. The README's human-evaluation table also compares KAgentSys with NoAgent, ReACT, and Auto-GPT; those figures are the project's reported evaluation results.

Key facts side by side with the most closely related agents.

Agent Source review Stars Updated Language Full support on
KwaiAgents This agent 0 · Major gaps ★ 1.2k 2y ago Python OpenAI API
AgenticSeek 40 · Major gaps ★ 27k 2d ago Python OpenAI API
Event Planning Multi-Agent System 62 · Some gaps ★ 115 5mo ago Python
Hermes Web UI 51 · Major gaps ★ 19k today Python Codex · Claude Code · OpenAI API · Claude API

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
0/ 100 5-point scale 0.0 / 5
Trust 0/29
Reliability 0/14
Adaptability 0/18
Convention 0/18
Effectiveness 0/13
Verifiability 0/8
Why each dimension lost points
Trust0 / 29 · 0.0/5

Insufficient evidence: the repository provides no documentation on permission management, user confirmation, data flow transparency, sensitive data handling, dependency security, external effects, rollback, or source attribution. All criteria lack evidence, hence 0.

Reliability0 / 14 · 0.0/5

Insufficient evidence: no documentation on self-consistency, dependency availability, or failure messages. All criteria lack evidence, hence 0.

Adaptability0 / 18 · 0.0/5

Insufficient evidence: no documentation on target audience, capability boundaries, trigger precision, or environment fit. All criteria lack evidence, hence 0.

Convention0 / 18 · 0.0/5

Insufficient evidence: no documentation on information architecture, installation notes, naming stability, examples/FAQ, known limitations, license, versioning/changelog, or maintenance responsibility. All criteria lack evidence, hence 0.

Effectiveness0 / 13 · 0.0/5

Insufficient evidence: no documentation on output usability, marginal value, or cost-benefit. All criteria lack evidence, hence 0.

Verifiability0 / 8 · 0.0/5

Insufficient evidence: no documentation on claim traceability, cross-source corroboration, or fact/inference separation. All criteria lack evidence, hence 0.

Risks and how to mitigate them
  • Not found in source: least-privilege scopingGrant only what the task needs: a dedicated account or read-only token, scoped to specific directories and repos.
  • Not found in source: confirmation before actingTurn on (or add) a confirmation step before it acts, and try it in a sandbox or test environment before real data.
  • Not found in source: data-flow disclosureWatch which external services it contacts (proxy or firewall logs) and keep sensitive data out until you know where it goes.
  • Not found in source: sensitive-data handlingUse dedicated, low-privilege, revocable API keys — never production credentials — and keep secrets out of logs.
  • Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
  • Not found in source: disclosed external effectsEstablish which external systems it writes to, sends to or changes, and verify with test accounts or repos before production.
  • Not found in source: rollback or recovery pathBack up first, or work on a git branch or snapshot, so its changes can be undone.
  • Not found in source: verifiable attributionInstall from the official repo or registry and check the publisher and URL to avoid look-alike packages.
  • The repository provides no documentation on security, permissions, or data handling; assess risks before use.
  • No dependency security audit or known vulnerability information is provided; dependency versions may be outdated.
  • No clear maintenance responsibility or update path is provided; the project may lack ongoing maintenance.
Evidence confidence: Low Reviewed Aug 09, 2026 Reviewed revision 3504ab8a00a3
See the full review method →

FAQ

What credentials are needed for GPT-3.5-turbo?
Set OPENAI_API_KEY. Weather-related queries additionally require WEATHER_API_KEY.
Does KAgentSys-Lite include memory?
No. The project explicitly states that the Lite system lacks memory mechanisms.
Can it run without a GPU?
Yes. The README documents a CPU-serving path using llama-cpp-python and the Qwen-7B-MAT GGUF model.
What should I check if browsing or search fails?
browse_website requires chromedriver. For repeated search failures, the README notes that duckduckgo_search may be inaccessible and suggests setting http_proxy.
View on GitHub ↗ Install ↓

Related agents