Architecture

How the Reactor thinks.

The quality of AI output depends on process design, not prompt engineering skill. The Reactor shifts control: away from prompting, toward designing cognitive recipes.

200+Methods
84+Recipes
63Workflows
3Model Slots
4Personas
6Patterns
Prompt Architecture

One stream.
Five composable layers.

Every API call builds a system prompt from five independent layers. The same recipe step produces fundamentally different results depending on persona, context, and method. Library of 200+ frameworks.

01

Persona

Cognitive stance — analytical, creative, adversarial, or curatorial

02

Method

One of 200+ thinking methods, injected as a structured instruction

03

Context

Pinned content + pruned history (full, last_step, or none)

04

Slot

Model selection — LOGIC, CREATIVE, or SEARCH per step

05

Output

Structured nodes, streamed in real-time with transparent reasoning

Multi-Model Orchestration

Three slots.
Heterogeneous composition.

Each recipe step declares the required cognitive slot. Analytical work goes to reasoning models, research to web models, synthesis to generative models — all in one workflow.

SLOT_LOGIC

Analytical Reasoning

Reasoning models (o1, o3-mini, Gemini Flash Thinking)

TRIZ analysis, contradiction detection, scoring, validation, adversarial audit

SLOT_CREATIVE

Ideation & Synthesis

Generative models (Claude Sonnet, GPT-4o, Llama 4)

Brainstorming, storytelling, concept synthesis, solution architecture, copywriting

SLOT_SEARCH

Evidence-Based Research

Web models (Perplexity Sonar, Gemini with grounding)

Patent research, competitive analysis, market research, state of the art

Orchestration Patterns

Six ways to orchestrate thinking.

From linear pipelines to autonomous, self-correcting swarms — every recipe is built on composable execution patterns.View all recipes

PATTERN A

Linear Pipeline

Sequential steps, each building on the previous one. Simple, deterministic, fast.

Step 1Step 2Step 3full ctxfull ctxfull ctx

e.g. TRIZ Express, Elevator Pitch

PATTERN B

Multi-Model Orchestrated

Different model slots per step — creative, analytical, and research models tackle the same problem from their respective strengths.

DraftCREATIVEResearchSEARCHAuditLOGICnonelast_stepfull

e.g. TRIZ v8, Pitch Architect

PATTERN C

User-Driven Branching

The engine pauses at a decision point. The user chooses a path, and the recipe branches accordingly.

PAUSEInputPath APath BOption 1Option 2

e.g. Dilemma Decoder, Branching Demo

PATTERN E

Self-Correcting Loop

Autonomous generate → verify → revise cycle. AI evaluates its own output and refines until quality standards are met.

GenerateAI CheckNextPassFail → Retry

e.g. Aletheia Engine, TRIZ v9 MAX

PATTERN F

Parallel Swarm

Multiple agents work simultaneously with different perspectives. Outputs are aggregated and synthesized.

SpawnLogicCreativeResearchΣSynth

e.g. MAD Engine, Ergodic Hive

HYBRIDS

Composable Patterns

Patterns can be freely combined. Real recipes are hybrids:

  • E+FAutonomous swarm — parallel agents + self-correction
  • C+FBranching swarm — user decision → specialized swarm
  • B+ESlot-based correction — creative generates, logic verifies
  • D+ECopilot / Autopilot — user switches between manual and autonomous
Transparency

Glass Box UI.

Every AI response is fully transparent. No black box. See what the model thought before it wrote — inspect structured data and debug raw output — all in real-time.

Agent labels show which specialist is active during multi-step recipes. The Cognitive Protocol reveals the model's internal reasoning process before a single word of output appears.

Rendered
Thinking
JSON
Raw
Rendered

Formatted output — clean, structured, actionable

Thinking

Native reasoning trace from the model's thought process

JSON

Structured node data — parseable, exportable

Raw

Unprocessed model output for debugging

Showcase

TRIZ v8 Orchestrator.

5 specialized agents. 3 model slots. 2 context modes. The flagship recipe demonstrates full orchestrator capacity.

#1System Analysis

Decompose the problem into components, functions, and contradictions.

SLOT_LOGICctx:full
#2Logic Engine

Apply the TRIZ contradiction matrix. Generate solution directions.

SLOT_LOGICctx:last_step
#3Patent Swarm

Research existing solutions, patents, and analogous domains.

SLOT_SEARCHctx:last_step
#4Solution Architect

Synthesize all insights into concrete, actionable concepts.

SLOT_CREATIVEctx:full
#5Adversarial Audit

Identify weaknesses, risks, and unintended consequences.

SLOT_LOGICctx:full

Context Pruning

Steps 2 & 3 use last_step — the model focuses on the distilled output, not the full history.

Slot Diversity

Analysis → reasoning models. Research → web models. Synthesis → creative models. No single model does everything.

Adversarial Close

Step 5 sees the full context but uses a reasoning model — maximum scrutiny on the entire proposal.

Scientific Foundation

Why cognitive architecture beats mega-prompts.

In March 2026, the Kimi team published the paper Attention Residuals — a mathematical proof of a problem that the Reactor already solves architecturally at the process control level.

Neural Level (Kimi)

PreNorm Dilution

In deep networks, residual connections accumulate. Essential early information gets diluted by noise from middle layers — deeper layers lose access to the original signals.

Application Level (Reactor)

Context Contamination

In standard chats, context accumulates message by message. By the time a model reaches step 6, the original problem is diluted in the noise of iterative intermediate steps.

Symmetry 01

Skip Connections → Epistemic Anchor

Kimi: Later layers skip accumulated noise and directly access early, clean layers (Attention Residuals).

Reactor: The Epistemic Anchor preserves extracted facts from step 1 via auto-pinning. Later agents with context_mode: last_step work only with the distilled output — the ground truth remains directly accessible via <pinned_context>.

Symmetry 02

Block Compression → Swarm Synthesis

Kimi: Layers are grouped into blocks and compressed into a single vector. Later layers see only the clean summary, not the raw individual steps.

Reactor: When parallel agents work (Pattern F), a synthesizer step compresses the outputs into a dense XML aggregate. Subsequent steps process only this node — not the individual agent responses.

Symmetry 03

Deep & Narrow → Microsteps

Kimi: Networks with Attention Residuals reach their optimum with deeper, narrower architectures — shallow networks stagnate.

Reactor: Many tightly focused steps (deep & narrow) beat few overloaded mega-prompts (shallow & wide). When the engine handles navigation and context pruning, the model can concentrate 100% of its parameters on pure transformation.

Intelligence in complex systems doesn't emerge from endlessly accumulating data, but from targeted noise reduction — regardless of whether the system is a neural network or a cognitive pipeline.

Source: Kimi / Moonshot AI, Attention Residuals (March 2026) · View benchmark results →

Methods think ahead. Think along. Think further.

200+ methods. 84+ recipes. Three model slots — ready to steer.

Launch Reactor