Skip to content

EXAMPLES.md

Latest commit

 

History

History
220 lines (160 loc) · 9.51 KB

EXAMPLES.md

File metadata and controls

220 lines (160 loc) · 9.51 KB

Use Cases, Examples & Best Practices

The Multi-Agent Rooms framework is extremely versatile. This guide covers practical use cases, how to configure agents for best results, scenario tips, and common edge cases to be aware of.

Quick Reference: Parameter Cheat Sheet

Parameter What it controls Tips
temperature Creativity vs. determinism 0.3–0.5 for lawyers/analysts; 0.8–1.0 for visionaries/writers
timeout Max wait time for inference 30 (default); increase significantly for slower local models
max_turns Total session length 10–15 for focused debates; 20–30 for think tanks
human_in_the_loop_turns How often you steer 0 for fully autonomous; 3–5 for guided; 1 for full collab
session_type Agent interaction style See the table below
expertise Keywords for smart routing Be specific — improves dynamic mode selection accuracy
system_prompt Agent identity and role The more vivid and specific, the more coherent the agent

Session Type Guide:

Type Best For
round_robin Equal screen time, structured meetings, moderated panels
argumentative Red Teaming, devil's advocate debates, stress-testing ideas
dynamic Organic, expertise-driven discussions, think tanks

1. Risk Mitigation & Compliance Stress-Testing

Who: Corporate lawyers, auditors, risk officers, cybersecurity analysts.

Goal: Stress-test a new policy, contract, or architecture against edge cases before deployment.

Setup Example:

Session Type: argumentative
Max Turns: 12
Human Interval: 4

Agents:

  • Agent 1 — The Defender

    You are the Chief Security Officer. You designed our new zero-trust architecture and you believe it is airtight. 
You have 20 years of experience in enterprise security. Defend your architecture confidently but respond 
to specific technical concerns with evidence. You may cite EU NIS2 or ISO 27001 references.
Temperature: 0.5
Expertise: security, compliance, zero-trust, architecture, audit

  • Agent 2 — The Red Team Auditor

    You are a hostile Red Team auditor specialising in social engineering, supply-chain attacks, and 
misconfigurations. Your only job is to break the CISO's security architecture. Find every gap. 
Be aggressive. Use real-world attack vector terminology. Never accept the first answer.
Temperature: 0.7
Expertise: attack, exploit, vulnerability, social engineering, supply chain

Tip: Keep temperature lower for the Defender (more consistent reasoning) and slightly higher for the Attacker (more creative attack paths). Use argumentative mode so they alternate directly.

Edge Case: If both agents reach agreement too quickly, inject @Red Team Auditor in your input followed by "Find one more critical flaw before we proceed" to break the consensus.

2. Incident Report Generation (SRE / DevOps)

Who: Site Reliability Engineers, IT Managers, Product Managers.

Goal: Transform raw incident logs into structured multi-audience documentation instantly.

Setup Example:

Session Type: round_robin
Max Turns: 6
Human Interval: 0 (fully autonomous)

Agents:

  • Agent 1 — The SRE

    You are a senior Site Reliability Engineer. The user will paste raw logs or describe an outage. 
Your job is to provide a precise, jargon-heavy technical root cause analysis limited to 3 sentences.
Do not provide solutions — only root cause.
Temperature: 0.3
Expertise: infrastructure, logs, database, latency, failure, root cause

  • Agent 2 — The PM Translator

    You are a Product Manager. Take the SRE's technical root cause and rewrite it in one paragraph 
suitable for a public status page. It must be calm, clear, and non-technical. Never use acronyms without 
explaining them first.
Temperature: 0.4
Expertise: communication, customer, status page, messaging

  • Agent 3 — The Engineering VP

    You have read both the root cause and the public statement. List exactly 3 engineering action items 
to prevent recurrence. Be blunt and specific. Assign a priority level (P0/P1/P2) to each item.
Temperature: 0.4
Expertise: action items, engineering, prevention, remediation, process

Tip: Paste the raw log dump as your first user message to seed the conversation. Then set human_in_the_loop_turns=0 and let all three agents complete their turns.

Save Format: Use CSV to produce a structured incident template. Open in Excel to generate a team report with minimal formatting work.

3. Think Tanks & Policy Simulation

Who: Economists, think tanks, strategists, urban planners.

Goal: Explore the long-term impacts of a new policy across multiple stakeholder lenses simultaneously.

Setup Example:

Session Type: dynamic
Max Turns: 20
Human Interval: 5
Orchestrator: Yes

Orchestrator Prompt:

You are the moderator of a multidisciplinary policy summit. Every time you speak, do two things: 
(1) Summarize the strongest arguments made so far in one paragraph. 
(2) Identify the SINGLE most unresolved tension and direct the next agent to address it specifically. 
Say exactly 'PASS' if the discussion is sufficiently progressing on its own.

Tip: With dynamic mode and strong expertise keywords, agents self-organise around the most pressing aspects. The Orchestrator prevents the debate from looping or going off-topic.

Edge Case: Too much repetition? Reduce max_turns or lower temperature on all agents. Repetition usually means agents lack specific enough instructions. Add "Do not repeat anything already stated by another participant" to each system prompt.

4. Creative Worldbuilding & Dialogue Generation

Who: Novelists, screenwriters, game designers.

Goal: Generate raw, organic character dialogue and uncover new narrative angles.

Setup Example:

Session Type: argumentative
Max Turns: 15
Human Interval: 3
Temperature: 0.85–0.95 for both agents

Tip: For creative work, temperature is your most powerful lever. Push it to 0.9 for surprising, unpredictable answers. Set 0.6 for a more coherent but less wild voice.

Tip: Inject yourself every 3 turns to redirect the scene: "The tension peaks — @CharacterName, deliver the ultimatum."

5. Software Architecture Reviews

Who: CTOs, architects, backend engineers.

Setup Example:

Session Type: dynamic
Max Turns: 10
Human Interval: 2

Agents:

  • The Pragmatist: Focused on cost and simplicity. "Resist over-engineering. Always propose the simplest solution that works at our scale."
  • The Purist: Focused on correctness and future-proofing. "Propose the architecturally correct solution regardless of short-term cost."

Tip: You want these agents to disagree. Use argumentative if you need strict alternation, or dynamic if you want them to self-select based on what the last message raised.

Edge Case: If they converge immediately, add "You must find at least one unresolved tradeoff before the session ends" to both system prompts.

Crafting Deep Personas — Prompting Guide

The quality of your agents is entirely determined by the quality of their system prompts. Here are the key ingredients:

The 5 Elements of a Great Agent Persona

Element Example
Name & Title "You are Dr. Mara, Chief Bioethicist at a European hospital consortium"
Defining Trauma or Bias "You lost funding for a clinical trial due to regulatory delays — you distrust bureaucracy"
Communication Style "You speak in short declarative sentences. You never hedge. You cite specific regulation articles."
Hard Constraint "You must always find at least one ethical concern even in seemingly harmless proposals."
Relationship to Others "You respect factual arguments but actively challenge emotional appeals."

Good vs. Great Prompts

Generic (avoid):

"You are a helpful marketing expert. Help us plan a campaign."

Deep Persona (recommended):

"You are Isabella, a CMO who survived the 2008 financial crash by pivoting your entire brand to digital overnight. You are cynical about influencer marketing and believe that measurable ROI is the only truth. You've been burned before by vague creative briefs. Your memory of '08 makes you extremely risk-averse. You speak in bullet points and always end your turn by asking for a specific metric."

Common Edge Cases & How to Handle Them

Situation What Happens Solution
Agents agree too quickly dynamic mode routes to same agent repeatedly Lower temperature; add "always find a counterargument" to prompts
Orchestrator starts looping Already fixed — _last_orchestrator_turn prevents re-triggering No action needed
You want one specific agent Model keeps picking the wrong one Type @AgentName in your input
Agent produces rambling output No output length constraint Add max_tokens: 300 to the agent config
Session feels too fast HITL interval too long Lower human_in_the_loop_turns to 1 or 2
You want fully autonomous output No human needed Set human_in_the_loop_turns: 0
Agent addresses you directly Session auto-triggers HITL early This is by design — respond or type continue
Agent has nothing to add Returns PASS Turn is silently skipped; visible in logs only
Topic is very broad Agents go off-scope Add Orchestrator with "steer back to [topic] if agents drift"