vibe: Tree-Search Research Engine for Agents

Where this fits

You're building an agent that must dig into a new scientific question — say, evaluating contradictory studies on a drug's mechanism. Naive retrieval loops either go shallow or get stuck in echo chambers. vibe replaces that with a structured tree search: your agent explores multiple hypotheses, logs every decision, and keeps going until a defensible conclusion emerges. It's for teams whose agents need to produce research-grade outputs, not just plausible-looking summaries.

Why agents benefit

• Agentic tree search (OTAE): Each node runs Observe-Think-Act-Evaluate, so the agent systematically branches hypotheses and prunes dead ends — no more linear, forgetful QA chains.
• Adversarial review built in: The skill forces counter-evidence search and a "red flag" checklist at each node, catching confirmation bias before it reaches the final report.
• Serendipity preserved: It doesn't just optimize for the obvious path; it injects inversion and collision brainstorm techniques to surface unexpected connections you'd otherwise miss.
• Rigorous tracking: Every step is journaled in a TreeNode, so you can trace why the agent chose one branch over another — essential for debugging and auditability.
• Auto-experiment protocol: When a hypothesis needs testing, the skill can design and execute an experiment loop, feeding results back into the tree without human handholding.

Practical scenarios

Literature review with conflicting evidence

Your agent must synthesize findings from 50+ papers on a contested topic. vibe lets it branch by methodology, sample size, and effect direction, then adversarially check each branch for missing counter-studies. The result is a balanced map of the evidence landscape, not a cherry-picked summary.

Novel hypothesis generation

You want your agent to propose new research directions from a set of known phenomena. With tree search, it can combine surprising pairs (inversion technique) and test feasibility of each combination, surfacing ideas a standard prompt would never reach.

Reproducible experiment design

For an agent that must run simulations, vibe's 5-stage experiment manager enforces planning, execution, validation, and reporting at every node. The DOI verification step ensures referenced protocols are real, preventing hallucinated citations from polluting the process.

Add it to your agent workflow

Install via:

npx skills add th3vib3coder/vibe-science --skill vibe

Then invoke it by describing your research goal, e.g., "Investigate the role of microRNAs in tau protein clearance, considering both inhibitory and enhancing pathways." The skill will return a tree of explored hypotheses, evidence logs, and a final synthesis with confidence scoring.

Compare before adopting

Look beyond the star count (16 at time of writing). Evaluate maintenance freshness — v4.5.0 updated recently with a codename and changelog, which signals active iteration. Compare against alternatives like plain tree-of-thought prompting or SAKI-like frameworks. Check the README's "WHY THIS SKILL EXISTS" section: it's opinionated and detailed, a good sign. Ensure the OTAE discipline fits your agent's tool-use patterns and that you're comfortable with the infinite-loop-until-discovery philosophy — it may need guardrails for bounded tasks.

Why it is worth tracking

Vibe teases serious ambition: lineage from Sakana AI-Scientist-v2 and a structured lineage of its own (v3.5 → v4.0 → v4.5). The adversarial review and serendipity features are differentiators few research agents offer. For developers building scientific reasoning agents, this skill is worth a trial run. Keep an eye on its GitHub issues and commit history — if it gains traction and community contributions, it could become a reference implementation for agentic research. Evaluate it when you need a research layer that will stand up to adversarial probing, not just answer nicely. For a full breakdown of its architecture and tree-search behavior, check the OpenAgentSkill listing.