Decision Guide
Which Architecture Solves Your Problem?
Start with the business challenge. We'll show you the architectures built for that exact scenario -- and when to choose one over another.
I need higher-quality AI outputs
Your AI generates content that requires heavy editing. Code has bugs. Marketing copy sounds generic. Reports miss key points. You need AI that produces polished, accurate output -- not rough drafts that create more work than they save.
Self-Refining AI
Reflection #01When to Choose
You need better output right now -- quality improvements within a single interaction. Best for one-off or varied tasks where each prompt is different.
Key Differentiator
Self-critique loop: generate, review, refine. Immediate improvement, no memory between sessions.
Continuously Learning AI
RLHF #15When to Choose
You perform the same type of task repeatedly and want quality to improve over time.
Key Differentiator
Critic-driven revision with persistent memory. Saves approved examples and references them for future tasks. Gets better the more you use it.
When to choose which: If you're writing one blog post, use Self-Refining AI. If you're running a content production pipeline that writes 50 emails a month and needs to learn your brand voice, use Continuously Learning AI.
| Self-Refining AI | Continuously Learning AI | |
|---|---|---|
| Improves within a session | Yes | Yes |
| Improves across sessions | No | Yes |
| Needs training examples | No | Builds them automatically |
| Best for | Ad hoc quality improvement | Repeated production tasks |
I need AI connected to live data and APIs
Your AI answers with outdated information. It can't check inventory, look up account details, or search your knowledge base. You need an agent that reaches into your live systems to give answers grounded in current facts.
Real-Time Data Access
Tool Use #02When to Choose
Simple queries where one or two tool calls suffice. 'What's the current stock price?' 'Look up order #12345.'
Key Differentiator
Single-step tool invocation. Fast, simple, reliable for straightforward lookups.
Adaptive Research Agent
ReAct #03When to Choose
Complex queries where each answer informs the next question. Multi-hop research where the full path can't be known upfront.
Key Differentiator
Reasons between steps -- adapts its strategy based on what it discovers. Chains multiple searches dynamically.
Self-Healing Pipeline
PEV #06When to Choose
Any data pipeline where the sources are unreliable -- APIs time out, data might be stale, or results need validation.
Key Differentiator
Adds verification after every step. If a tool call fails or returns bad data, it replans with an alternative strategy automatically.
When to choose which: One API call? Tool Use. A chain of dependent searches? ReAct. Any of the above but the APIs are flaky? Wrap it in PEV.
| Real-Time Data Access | Adaptive Research Agent | Self-Healing Pipeline | |
|---|---|---|---|
| Number of tool calls | 1-2 | Many (chained) | Any (with verification) |
| Adapts between steps | No | Yes | Yes + replans on failure |
| Error recovery | None | None | Intelligent replan |
| Best for | Simple lookups | Investigative research | Unreliable data sources |
I need to automate complex, multi-step workflows
You have processes that involve multiple steps, potentially with different specialists or decision points. Manual coordination is slow, error-prone, and doesn't scale. You need AI that breaks down complex work, delegates it, and assembles the result.
Structured Workflow Engine
Planning #04When to Choose
The sequence of steps is predictable and can be determined upfront. Batch operations, report generation, compliance checklists.
Key Differentiator
Plans all steps first, then executes them in order. Transparent, predictable, efficient for known workflows.
Specialist Team AI
Multi-Agent #05When to Choose
The work requires different types of expertise -- and each specialist can work independently on their piece.
Key Differentiator
Multiple expert agents each handle their domain. A manager synthesizes the pieces into a cohesive deliverable.
Dynamic Decision Router
Blackboard #07When to Choose
The next step depends on what was discovered in the previous step. Conditional, branching workflows that can't be planned upfront.
Key Differentiator
A controller dynamically routes to the right specialist based on accumulated context. Skips irrelevant steps entirely.
When to choose which: If you can write the steps on a whiteboard before starting, use Planning. If the work needs multiple domain experts, use Multi-Agent. If 'it depends' is the answer to what comes next, use Blackboard.
| Structured Workflow Engine | Specialist Team AI | Dynamic Decision Router | |
|---|---|---|---|
| Steps determined | Upfront | Upfront (by specialty) | Dynamically (at runtime) |
| Multiple experts | No | Yes | Yes |
| Conditional branching | No | No | Yes |
| Best for | Predictable pipelines | Domain-expert analysis | Conditional workflows |
I need AI that remembers context across conversations
Your AI treats every conversation like it's meeting the user for the first time. Customers repeat themselves. Preferences are lost. Relationships can't develop. You need AI with persistent memory that builds understanding over time.
Persistent Memory AI
Episodic + Semantic #08When to Choose
The AI needs to remember what happened (conversations, preferences, events) and what it learned (facts, entity relationships).
Key Differentiator
Dual memory: episodic (past interactions as vectors) + semantic (extracted knowledge as entities). Answers both 'What did we discuss?' and 'What do I know about this person?'
Knowledge Graph Intelligence
Graph Memory #12When to Choose
The AI needs to reason across complex relationships -- who owns what, who reports to whom, multi-hop relational queries.
Key Differentiator
Structured graph database (Neo4j) storing entities and relationships. Can traverse chains: company, acquisition, subsidiary, product, competitor.
When to choose which: If you need to remember conversations and preferences, use Persistent Memory. If you need to reason across complex entity relationships (like an org chart or supply chain), use Knowledge Graph Intelligence. Many deployments use both.
| Persistent Memory AI | Knowledge Graph Intelligence | |
|---|---|---|
| Memory type | Conversations + extracted facts | Entities + relationships (graph) |
| Query style | What did they tell me before? | Who is connected to whom? |
| Best for | Personalized assistants | Relational reasoning (fraud, supply chain) |
| Infrastructure | FAISS vector store | Neo4j graph database |
I need to explore multiple options before deciding
Some decisions are too important for a single AI perspective. You need to evaluate multiple approaches, simulate outcomes, or get independent assessments before committing. The cost of a wrong decision outweighs the cost of thorough analysis.
Systematic Solution Finder
Tree of Thoughts #09When to Choose
The problem has a discrete set of possible solutions with clear validity constraints. Puzzles, scheduling, configuration search.
Key Differentiator
Explores a tree of possible solutions. Branches out, prunes invalid paths, finds the correct answer through systematic search.
Multi-Perspective Analyst
Ensemble #13When to Choose
The decision benefits from diverse viewpoints -- different analysts may weigh factors differently, and you want a balanced synthesis.
Key Differentiator
Multiple independent analysts assess the same problem in parallel. A senior synthesizer weighs agreements and disagreements.
Risk Simulation Engine
Mental Loop #10When to Choose
The decision has real-world consequences and you want to simulate outcomes before committing.
Key Differentiator
Proposes a strategy, forks into multiple simulated environments, observes the outcome distribution, then calibrates the final action.
When to choose which: Looking for the provably correct answer to a constraint problem? Tree of Thoughts. Want diverse expert opinions synthesized into one recommendation? Ensemble. Need to simulate 'what happens if we do X' before acting? Mental Loop.
| Systematic Solution Finder | Multi-Perspective Analyst | Risk Simulation Engine | |
|---|---|---|---|
| Approach | Tree search with pruning | Parallel diverse analysts | Propose, simulate, refine |
| Best for | Constraint satisfaction | Balanced judgment calls | High-stakes what-if decisions |
| Output type | The correct answer | Weighted recommendation | Calibrated action with risk profile |
I need safety and human oversight for AI actions
You want to use AI for tasks with real consequences -- publishing content, executing trades, making medical assessments, deploying infrastructure changes. But you can't afford mistakes. You need AI that either waits for human approval or knows when to stop and ask for help.
Human Approval Gateway
Dry-Run #14When to Choose
Every action should be previewed by a human before execution. Content publishing, financial transactions, infrastructure changes.
Key Differentiator
The agent proposes an action, shows a full preview, and waits for explicit human approval before executing. Rejected actions never go live.
Self-Aware Safety Agent
Metacognitive #17When to Choose
The agent handles a mix of routine and high-risk queries. It should act autonomously on easy tasks but automatically escalate when uncertain.
Key Differentiator
Maintains an explicit self-model of its own capabilities. Scores its confidence on every query. High confidence: acts. Low confidence: escalates to a human.
When to choose which: If every action needs human approval (like publishing or trading), use Dry-Run. If most actions are routine but some are dangerous (like medical triage), use Metacognitive -- it decides when to escalate, so humans aren't bottlenecked on easy tasks.
| Human Approval Gateway | Self-Aware Safety Agent | |
|---|---|---|
| Human involvement | Every action (explicit approval) | Only when uncertain (auto-escalation) |
| Bottleneck risk | Yes -- human must approve everything | No -- routine tasks proceed autonomously |
| Best for | High-consequence actions (publish, trade) | Mixed-risk domains (medical, legal) |
| Safety approach | Show me before you do it | Know when you don't know |
I need a modular, extensible AI system
You're building a platform that needs to handle many types of requests -- and you want to add new capabilities without rearchitecting everything. You need a system where new specialists can be plugged in and requests are automatically routed to the right handler.
Intelligent Task Router
Meta-Controller #11When to Choose
You have a finite number of distinct task types and need one front door that dispatches to the right specialist.
Key Differentiator
A single controller classifies the incoming request and routes it to the appropriate specialist agent. Adding a new capability = adding a new specialist.
Emergent Coordination System
Cellular Automata #16When to Choose
You have thousands of simple agents that need to coordinate at scale -- warehouse robots, sensor networks, traffic systems -- without a central controller.
Key Differentiator
No central planner. Each agent follows simple local rules. Optimal behavior emerges from the interactions of many simple units. Scales to massive numbers.
When to choose which: If you need a smart front door for a handful of specialist agents, use Meta-Controller. If you need thousands of simple agents coordinating through local interactions (robotics, logistics, sensor networks), use Cellular Automata.
| Intelligent Task Router | Emergent Coordination System | |
|---|---|---|
| Central controller | Yes (classifier) | No (purely local rules) |
| Number of agents | 3-10 specialists | Hundreds to thousands |
| Adding capabilities | Add a new specialist agent | Add new cell rules |
| Best for | Multi-purpose chatbots, support platforms | Warehouse robotics, swarm coordination |
What About Combining Architectures?
Most production deployments combine 2-4 architectures. Common patterns: ReAct + PEV: Multi-hop research with verification at every step. Multi-Agent + Ensemble: Specialist analysis with multi-perspective synthesis. Metacognitive + Dry-Run: Self-aware escalation for risky tasks, human approval for critical ones. Episodic Memory + Meta-Controller: A front door that remembers every user's preferences and history. Planning + PEV + Dry-Run: Structured workflow with verification and human approval gate.