Multi-Agent Technical
Decision System
A structured reasoning engine that helps engineers make high-stakes technical decisions through transparent, multi-agent analysis with explicit disagreement handling and cost awareness.
Explicit Disagreement
Unlike single-LLM systems, we surface conflicts between agents. The Disagreement Detector identifies where specialists contradict each other.
Cost Transparency
Every decision iteration costs ~$0.02. We track tokens, cost per agent, and total spend in real-time. No hidden API charges.
Critic & Gate
A dedicated Critic Agent challenges weak reasoning. The Gate Agent enforces confidence thresholds and can force deferral if uncertain.
Submit Technical Decision
Multiple agents will reason independently and disagree constructively
Agent Execution Graph
Click on Agent to see input and outputPlanner
Decomposes
Systems
ML/AI
Cost
Product
Detector
Finds conflicts
Critic
Challenges
Synthesizer
Resolves
Gate
Validates
Live Agent Reasoning
Updates as agents completeSubmit a decision prompt to observe agent reasoning
Detected Disagreement
Critic Analysis
Select risks to acceptFinal Synthesis
Iteration 2 Feedback
Review the critic's issues above and select which risks you're willing to accept.
System Architecture
Designed to avoid single-LLM hallucination through structured multi-agent reasoning with explicit disagreement handling.
1. Planner
Decomposes into sub-questions
2. Specialists
Parallel independent evaluation
3. Detector
Finds conflicts explicitly
4. Critic
Challenges assumptions
5. Synthesizer
Unifies recommendation
6. Gate
Final validation
Narrow Roles
Each agent has a strictly defined scope. The Systems Agent cannot evaluate ML feasibility. The Cost Agent cannot assess user experience. This constraint prevents single-LLM scope creep.
Explicit Disagreement
The Disagreement Detector forces conflicts into the open. When Systems recommends batch and Product demands real-time, this conflict is surfaced, categorized by severity, and must be resolved.
Schema Validation
All agent outputs are validated against strict schemas. Confidence scores must be 0-1. Recommendations must be from an enum. Rationale must be a string array. No free-form text blobs.
Agent Specifications
Each agent is a specialized reasoning unit with constrained prompts and explicit outputs.
Planner Agent
GPT-5-miniDecomposes ambiguous technical decisions into structured sub-questions for specialist agents. Forces clarity before reasoning begins.
Systems Agent
GPT-5-miniEvaluates infrastructure, scalability, latency, and operational complexity. Focuses on batch vs online, reliability, and ops overhead.
ML/AI Agent
GPT-5-miniAssesses model complexity, training vs inference costs, data requirements, and MLOps overhead. Evaluates technical feasibility of ML approaches.
Cost Agent
GPT-5-miniAnalyzes cloud costs, model inference expenses, and long-term scalability. Evaluates cost-performance trade-offs and budget impact.
Product Agent
GPT-5-miniEvaluates user experience, market fit, feature velocity, and business alignment. Considers latency tolerance and user expectations.
Disagreement Detector
GPT-5-miniIdentifies conflicts between specialist recommendations. Surfaces disagreements explicitly with severity ratings and blocking status.
Critic Agent
GPT-5-miniChallenges assumptions and identifies weak reasoning. Does not propose solutions—only attacks blind spots to force robustness.
Synthesizer Agent
GPT-5.1Integrates all perspectives into a final recommendation with confidence score, rationale, trade-offs, and unresolved risks.
Gate Agent
Rule BasedValidates decision quality against thresholds. Enforces minimum confidence, checks for unresolved blocking conflicts, manages approval tiers.
Cost Model & Infrastructure
Transparent pricing and serverless deployment on Google Cloud Run.
Cost Per Iteration
Iteration 2 (if needed) costs additional $0.02
Hosting Architecture
Google Cloud Run
Serverless container platform with automatic scaling, WebSocket support, and pay-per-use pricing. Single-digit USD/month for idle instance.
FastAPI + Uvicorn
High-performance Python backend with native WebSocket support for real-time agent execution streaming.
Docker Container
Reproducible deployments with locked dependency versions. Scales from zero to thousands of requests instantly.
Why Cost Transparency Matters
Budget Awareness
Users understand that AI decisions consume real resources. This prevents frivolous queries and encourages thoughtful decision formulation.
Iteration Trade-offs
Iteration 2 costs another $0.02. Users must consciously decide if resolving uncertainty is worth the additional cost, mirroring real engineering trade-offs.
Model Selection
Specialists use GPT-5-mini for cost efficiency. The Synthesizer uses GPT-5.1 for higher reasoning quality. This tiered approach optimizes cost/quality.
Open Source
Full backend implementation, agent logic, and deployment configurations available on GitHub.