lauraflorentin Exploration & Discovery
In the Exploration pattern, the goal is not to execute a known task, but to find new information or solutions. Agents act as scientists or researchers: they formulate a hypothesis, test it (by searching, coding, or simulating), analyze the results, and iterate. This allows for genuine novelty and discovery.
When to Use
- Literature Review: "Survey the field of Quantum Computing and identify gaps."
- Idea Generation: "Brainstorm 50 potential names for this product and check domain availability."
- Scientific Discovery: Analyzing large datasets to find correlations.
- Market Research: Exploring competitor websites to map out their feature sets.
Use Cases
- Agent Laboratory: A team of agents (Professor, Postdoc, Reviewer) writing a research paper.
- Creative Studio: Agents collaborating to write a screenplay or design a game.
- Scenario Planning: Simulating how a stock portfolio would perform under various economic conditions.
Implementation Pattern
def exploration_loop(topic):
knowledge_base = []
# Phase 1: Hypothesis Generation
hypotheses = brainstorming_agent.run(f"Generate ideas about {topic}")
for hypothesis in hypotheses:
# Phase 2: Experiment / Research
# Agent autonomously decides search queries or code to run
evidence = researcher_agent.run(f"Test this hypothesis: {hypothesis}")
# Phase 3: Analysis
conclusion = analyst_agent.run(
prompt="Does the evidence support the hypothesis?",
input={"hypothesis": hypothesis, "evidence": evidence}
)
knowledge_base.append(conclusion)
# Phase 4: Synthesis
return writer_agent.run("Write a report based on these conclusions", input=knowledge_base)
Examples
Input: "Survey the latest research on LLM memory architectures and identify the top 3 open problems."
What the agent does:
- Generates 5 hypotheses about unsolved memory challenges
- Searches arXiv for recent papers on each hypothesis
- Cross-references citation counts to rank relevance
- Synthesizes findings into a structured research brief
Output: A 2-page research brief with identified gaps and recommended next steps.
Input: "Explore competitor pricing pages and build a comparison matrix."
Output: A structured table comparing plans, features, and pricing across 8 competitors, with a gap analysis highlighting differentiation opportunities.
Troubleshooting
| Problem | Cause | Fix |
|---|---|---|
| Agent explores indefinitely | Missing stopping criteria | Set explicit iteration limit (max_iterations=10) and convergence condition |
| All hypotheses converge to same answer | Insufficient diversity in brainstorming | Add temperature variation or inject contrarian perspectives |
| Research results are outdated | Search tool returning cached content | Specify date filter (after:2024) in search queries |
| Agent loses track of prior findings | Context window exceeded | Use an external knowledge store; summarize findings at each iteration |
| Exploration produces no actionable output | No synthesis step | Always include a final writer_agent pass to consolidate findings |