---

name: webagent-family description: "Route DeepResearch/WebAgent family tasks to the right variant, prerequisites, blockers, and sibling sub-skills." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

WebAgent Family Router

Use this sub-skill when a user asks for help choosing, comparing, adapting, or troubleshooting one of the DeepResearch, WebAgent, or Agent family variants and does not already name the exact workflow. This is a router and operating-memory skill: it preserves the high-value family map, prerequisites, and workflow-selection knowledge in the bundled references so future agents do not need the original repository documentation.

Immediate routing rules

1. If the request is about Tongyi DeepResearch ReAct inference, .env setup, local vLLM serving, OpenRouter/OpenAI-compatible inference, ReAct input data, tools, or rollout files, route to sibling ../react-inference/SKILL.md.
2. If the request is about official judging, benchmark metric mechanics, prediction-rollout validation, Pass@k summaries, HLE/deep-search evaluation scripts, or LLM-as-judge costs, route to sibling ../benchmark-evaluation/SKILL.md.
3. Otherwise choose family variants from the task signals below, then read the bundled references in this sub-skill before proposing a run plan.

Fast task-signal chooser

User task signal	Prefer these family routes	Why	Common blockers
ReAct inference, search/visit agent demo, long-horizon QA with tools	DeepResearch via react-inference; WebDancer; WebSailor	Root DeepResearch and early WebAgent projects expose ReAct-style Search/Visit inference patterns.	Model weights, GPUs or hosted API, search/visit credentials, output/data schema.
Post-training, agentic data construction, SFT/RL, synthetic QA	WebDancer; WebSailor; WebSailor-V2; WebShaper; WebLeaper; AgentFounder; AgentScaler	These projects are primarily training/data/scaling methods rather than simple inference wrappers.	Some assets are paper-only or unreleased; training requires large data, model weights, and RL infrastructure.
Multimodal visual web search or VQA	WebWatcher	WebWatcher adds image search, visual reasoning, OCR/image benchmarks, and code interpreter tooling.	Trained model, summary model, large image archives, image/text search credentials, optional OSS credentials.
Open-ended deep report writing, citations, dynamic outline	WebWeaver; conceptual WebResearcher	WebWeaver has planner/writer code with dynamic outline and memory-grounded report synthesis; WebResearcher describes iterative report-memory rounds.	4x80G summary-model serving for WebWeaver, planner/writer APIs, scraper/search credentials.
Context summarization, restartable long-horizon search, proactive compression	WebResummer/ReSum; AgentFold; conceptual WebResearcher	ReSum periodically compresses trajectories; AgentFold proactively compresses previous steps; WebResearcher uses a running report as central memory.	Summary models/tool checkpoints, vLLM services, search/visit credentials, code placeholders.
Efficient/entity-intensive information seeking	WebLeaper; WebShaper; sometimes WebSailor-V2	WebLeaper targets high-density entity retrieval and efficiency rewards; WebShaper formalizes IS task synthesis.	Mostly data/training guidance; runnable inference code may be absent in this checkout.
Web traversal, RAG baselines, WebWalkerQA	WebWalker; sibling benchmark-evaluation for official metric mechanics	WebWalker contributes the WebWalkerQA benchmark, Streamlit demo, RAG baseline, and crawl4ai/Qwen-Agent stack.	crawl4ai setup, provider API keys, dataset acquisition, judge API.
Test-time scaling, rollout convergence, answer fusion	ParallelMuse; conceptual WebResearcher; AgentFold	ParallelMuse aggregates multiple rollout reports; WebResearcher documents last-k-fusion; AgentFold can produce long rollouts with compressed context.	Multiple rollouts, local OpenAI-compatible model endpoint, high context length, placeholders/source rough edges.
Nested browser-use with click/fill	NestBrowse	NestBrowse adds MCP/browser actions beyond Search/Visit: visit, click, fill, and nested page state.	Browser MCP server, local model endpoint, tokenizer path, data/results setup.
Continual pretraining or environment scaling	AgentFounder; AgentScaler	AgentFounder is Agentic CPT; AgentScaler is simulated environment/function-calling scaling.	Public docs mostly describe methodology; no full runnable training code in the inspected tree.

Required operating procedure

1. Capture the user’s intent in one sentence and identify the strongest task signals.

2. Run the safe local chooser when useful:

   python scripts/choose_webagent_variant.py "deep report with citations and dynamic outline"
   

3. Read references/family-map.md for variant capabilities and evidence status.

4. Read references/family-workflows.md before recommending commands, models, dependencies, GPUs, or API services.

5. Read references/troubleshooting.md when a requested family workflow is blocked by missing checkpoints/data, credentials, large archives, generated/vendored code, or path confusion.

6. Reroute back to react-inference or benchmark-evaluation whenever the user’s request is really root DeepResearch inference or official metric/evaluation mechanics.

Non-goals and safety boundaries

• Do not reproduce detailed root ReAct configuration here; use react-inference.
• Do not reproduce official benchmark metric mechanics here; use benchmark-evaluation.
• Do not tell future agents to open the original README tree. Distilled facts are in this sub-skill’s references.
• Do not claim paper-only or placeholder code is immediately runnable. The references distinguish runnable code, data/model notes, and blockers.
• Do not include local checkout paths, private environment prefixes, or secret values in user-facing plans.