Eight AI Agents Debate 2+2 in a 97-Step Workflow Built With YAML

I built 97 blocks of YAML, 8 AI specialists, and 18 phases of deliberation to answer a question every five-year-old knows: what is 2 + 2? It took the system about four minutes to reach a unanimous verdict: 4. Along the way, a mathematician invoked Peano axioms, a philosopher questioned whether numbers exist, a poet called four "the universe's quietest poem," and a child said "it's 4, everybody knows that." The most interesting character had only 72% confidence. This is how I built Rein, an open-source orchestrator for multi-agent AI workflows, and why I tested it on the most over-engineered arithmetic problem ever created. Rein The problem I have a dozen AI agents on my server. Research agents, code reviewers, content writers, critics. Each one works fine alone. Getting them to work together is the problem. One massive prompt with all the instructions works for simple tasks. For anything with more than two steps, the model forgets the beginning by the time it reaches the end, and when something fails you re-run everything from scratch. I needed a way to split work into steps, hand each step to a different specialist, and have results flow automatically from one to the next. The same structure that debates 2+2 can debate "should we use Kafka or RabbitMQ" -- the patterns are identical, only the question changes. Eight specialists walk into a debate Before the deliberation, I verified the engine with a 12-block flow that adds 2+2 using Python scripts -- no AI, just parallel validation, compute, verify, conditional branch. That was the unit test. The deliberation is the integration test. I defined 8 AI specialists, each with a different lens on the world: Specialist Perspective What they bring Mathematician Formal proof Peano axioms, logical certainty Philosopher Ontological "Do numbers even exist?" Physicist Empirical Two stones + two stones = four stones Child Naive common sense "It's 4, everybody knows that" Skeptic Challenges everything Modular arithmetic, Godel, doubt Historian 5000 years of consensus Babylonians knew this too Poet Aesthetic truth Beauty in numerical harmony Engineer Practical precision 4, +/- 0.01 tolerance Specialist Perspective What they bring Mathematician Formal proof Peano axioms, logical certainty Philosopher Ontological "Do numbers even exist?" Physicist Empirical Two stones + two stones = four stones Child Naive common sense "It's 4, everybody knows that" Skeptic Challenges everything Modular arithmetic, Godel, doubt Historian 5000 years of consensus Babylonians knew this too Poet Aesthetic truth Beauty in numerical harmony Engineer Practical precision 4, +/- 0.01 tolerance Specialist Perspective What they bring Specialist Specialist Perspective Perspective What they bring What they bring Mathematician Formal proof Peano axioms, logical certainty Mathematician Mathematician Formal proof Formal proof Peano axioms, logical certainty Peano axioms, logical certainty Philosopher Ontological "Do numbers even exist?" Philosopher Philosopher Ontological Ontological "Do numbers even exist?" "Do numbers even exist?" Physicist Empirical Two stones + two stones = four stones Physicist Physicist Empirical Empirical Two stones + two stones = four stones Two stones + two stones = four stones Child Naive common sense "It's 4, everybody knows that" Child Child Naive common sense Naive common sense "It's 4, everybody knows that" "It's 4, everybody knows that" Skeptic Challenges everything Modular arithmetic, Godel, doubt Skeptic Skeptic Challenges everything Challenges everything Modular arithmetic, Godel, doubt Modular arithmetic, Godel, doubt Historian 5000 years of consensus Babylonians knew this too Historian Historian 5000 years of consensus 5000 years of consensus Babylonians knew this too Babylonians knew this too Poet Aesthetic truth Beauty in numerical harmony Poet Poet Aesthetic truth Aesthetic truth Beauty in numerical harmony Beauty in numerical harmony Engineer Practical precision 4, +/- 0.01 tolerance Engineer Engineer Practical precision Practical precision 4, +/- 0.01 tolerance 4, +/- 0.01 tolerance Each specialist is a Markdown file with a system prompt. About 10 lines. The team YAML enforces brevity: "respond in 1-2 sentences, output JSON." 97 blocks, 18 phases The workflow YAML is 500+ lines. Here's the structure: Phase 1: OPENING STATEMENTS ............ 8 blocks (parallel) Phase 2: CROSS-REVIEW LEFT ............. 8 blocks (parallel) Phase 3: CROSS-REVIEW RIGHT ............ 8 blocks (parallel) Phase 4: REVISED POSITIONS ............. 8 blocks (parallel) Phase 5: STRUCTURED DEBATES ............ 8 blocks (4 pairs) Phase 6: DEBATE JUDGES ................. 4 blocks (parallel) Phase 7: POST-DEBATE POSITIONS ......... 8 blocks (parallel) Phase 8: SKEPTIC CHALLENGES ............ 7 blocks (parallel) Phase 9: RESPONSES TO SKEPTIC .......... 7 blocks (parallel) Phase 10: GROUP SYNTHESIS ............... 2 blocks (parallel) Phase 11: GROUP DEBATE .................. 2 blocks (sequential) Phase 12: MODERATOR SYNTHESIS ........... 1 block Phase 13: FINAL VOTE .................... 8 blocks (parallel) Phase 14: VOTE COUNT + CONSENSUS ........ 2 blocks (sequential) Phase 15: EMERGENCY ROUND ............... 4 blocks (conditional) Phase 16: EMERGENCY VOTE ................ 8 blocks (conditional) Phase 17: FORMAT ANSWER ................. 1 block Phase 18: QUALITY + DELIVER ............. 3 blocks (sequential) Phase 1: OPENING STATEMENTS ............ 8 blocks (parallel) Phase 2: CROSS-REVIEW LEFT ............. 8 blocks (parallel) Phase 3: CROSS-REVIEW RIGHT ............ 8 blocks (parallel) Phase 4: REVISED POSITIONS ............. 8 blocks (parallel) Phase 5: STRUCTURED DEBATES ............ 8 blocks (4 pairs) Phase 6: DEBATE JUDGES ................. 4 blocks (parallel) Phase 7: POST-DEBATE POSITIONS ......... 8 blocks (parallel) Phase 8: SKEPTIC CHALLENGES ............ 7 blocks (parallel) Phase 9: RESPONSES TO SKEPTIC .......... 7 blocks (parallel) Phase 10: GROUP SYNTHESIS ............... 2 blocks (parallel) Phase 11: GROUP DEBATE .................. 2 blocks (sequential) Phase 12: MODERATOR SYNTHESIS ........... 1 block Phase 13: FINAL VOTE .................... 8 blocks (parallel) Phase 14: VOTE COUNT + CONSENSUS ........ 2 blocks (sequential) Phase 15: EMERGENCY ROUND ............... 4 blocks (conditional) Phase 16: EMERGENCY VOTE ................ 8 blocks (conditional) Phase 17: FORMAT ANSWER ................. 1 block Phase 18: QUALITY + DELIVER ............. 3 blocks (sequential) Every phase runs in parallel where possible. The dependency graph looks like a circuit board: wide parallel bands feeding into narrow synthesis points, then fanning out again. The conditional branch at Phase 14 is the interesting part. If 75% of specialists agree, skip to formatting. If not, trigger emergency rounds. In this run, consensus was 100%, so the emergency path never fired. But it's there. What actually happened Opening statements Eight specialists answered simultaneously. Most said "4" without hesitation. The Mathematician cited Peano axioms. The Engineer said "4, fundamental arithmetic identity." The Historian traced it to Babylonian mathematics circa 3000 BCE. Two stood out. The Poet: "Two and two make four -- a truth as old as counting stars, where symmetry finds its most perfect mirror in the doubling of the pair. It is the first lesson of arithmetic, and perhaps the universe's quietest poem." "Two and two make four -- a truth as old as counting stars, where symmetry finds its most perfect mirror in the doubling of the pair. It is the first lesson of arithmetic, and perhaps the universe's quietest poem." And the Skeptic, at 72% confidence: "While 2+2=4 is treated as self-evident, this relies entirely on accepting the axioms of Peano arithmetic. In modular arithmetic mod 4, 2+2=0. The answer is framework-dependent, not absolute truth." "While 2+2=4 is treated as self-evident, this relies entirely on accepting the axioms of Peano arithmetic. In modular arithmetic mod 4, 2+2=0. The answer is framework-dependent, not absolute truth." Everyone else was at 97-100%. The Skeptic opened at 72%. This gap drives the entire deliberation. The debates Two rounds of cross-review planted the seeds -- the Philosopher called the Mathematician's axiomatic dependence "a significant philosophical concession," the Physicist dismissed modular arithmetic as "contextually irrelevant." Then four structured debates ran in parallel. Mathematics vs. Philosophy. The Mathematician argued that Peano axioms aren't arbitrary -- they capture our actual concept of natural numbers. The Philosopher rebutted: logical necessity within a system isn't the same as necessity full stop. "If these axioms, then this result" is not the same as "this result, period." Mathematics vs. Philosophy. They partially converged. Both agreed on 4; disagreed on whether it's a mind-independent truth or a consequence of a chosen framework. Physics vs. Skepticism. This was the turning point. The Physicist invoked empirical evidence: two stones plus two stones always makes four stones. Two photons plus two photons, same thing. Every measurement confirms it. Physics vs. Skepticism. The Skeptic conceded: "Physical reality does provide a non-circular grounding for why standard arithmetic is the privileged context, making '4' not merely conventional but empirically anchored." "Physical reality does provide a non-circular grounding for why standard arithmetic is the privileged context, making '4' not merely conventional but empirically anchored." The circularity -- "Peano axioms formalize standard arithmetic, which is standard because Peano axioms formalize it" -- was broken by pointing outside mathematics entirely. Physics grounds the math. Poetry vs. Engineering. The Poet argued mathematics without meaning is "a skeleton without flesh." The Engineer countered that wonder validates the arithmetic, not the other way around. Then the Engineer surprised everyone: Poetry vs. Engineering. "Precision alone is insufficient for full understanding. Context, meaning, and human experience are legitimate dimensions of knowledge, just orthogonal to the arithmetic fact itself." "Precision alone is insufficient for full understanding. Context, meaning, and human experience are legitimate dimensions of knowledge, just orthogonal to the arithmetic fact itself." The Skeptic's challenge round Phase 8 was the Skeptic's chance to attack everyone individually. Seven targeted challenges, deployed in parallel. To the Mathematician, the Skeptic invoked Godel: "If the consistency of Peano arithmetic cannot be proven within Peano arithmetic itself, your confidence rests on an unprovable meta-assumption." To the Child: "You're importing a specific foundational framework as if it were neutral." To the Poet, the sharpest attack: "If 2+2=4 is merely system-relative, then so is the beauty you invoke to defend it. Your certainty is self-undermining." All seven stood firm. The Mathematician pointed out that Godel's limitations apply equally to every alternative system -- there's no better option to switch to. The Child cut through the philosophy: "The question '2+2=?' implicitly operates within standard arithmetic. Acknowledging that other systems exist doesn't undermine the answer within the default context." The Poet made the most elegant counter-move: "A poet who says 'all roses are red within this garden' need not pretend to stand outside all gardens to say so with conviction." "A poet who says 'all roses are red within this garden' need not pretend to stand outside all gardens to say so with conviction." The Skeptic's arc Opening: 72%. "The answer is framework-dependent, not absolute truth." Opening: 72%. After the Physics debate: ~85%. The moment two stones plus two stones became the argument, the Skeptic's circularity attack lost its teeth. He conceded that standard arithmetic is "empirically anchored" -- not just a convention we happen to use. After the Physics debate: ~85%. Post-debate reflection: 88%. "The philosophical nuance about axiomatic choice is valid but doesn't undermine the practical answer." The hedging shrank. The confidence grew. Post-debate reflection: 88%. Final vote: 100%. "2+2=4 is a fundamental arithmetic fact, universally true in standard integer arithmetic." Final vote: 100%. The Skeptic didn't cave to pressure. He engaged every argument, found the best counterarguments, and changed his mind when the evidence warranted it. His journey from principled doubt to informed certainty is the intellectual backbone of the entire deliberation. Final vote Specialist Answer Confidence Mathematician 4 100% Philosopher 4 100% Physicist 4 100% Child 4 100% Skeptic 4 100% Historian 4 100% Poet 4 100% Engineer 4 100% Specialist Answer Confidence Mathematician 4 100% Philosopher 4 100% Physicist 4 100% Child 4 100% Skeptic 4 100% Historian 4 100% Poet 4 100% Engineer 4 100% Specialist Answer Confidence Specialist Specialist Answer Answer Confidence Confidence Mathematician 4 100% Mathematician Mathematician 4 4 100% 100% Philosopher 4 100% Philosopher Philosopher 4 4 100% 100% Physicist 4 100% Physicist Physicist 4 4 100% 100% Child 4 100% Child Child 4 4 100% 100% Skeptic 4 100% Skeptic Skeptic 4 4 100% 100% Historian 4 100% Historian Historian 4 4 100% 100% Poet 4 100% Poet Poet 4 4 100% 100% Engineer 4 100% Engineer Engineer 4 4 100% 100% Unanimous. 8/8, 100% confidence. The consensus check passed instantly. Emergency rounds never triggered. Transcript highlights These are real outputs from a real run. 89 of 97 blocks completed. Here are 10 moments from the deliberation: Child, opening statement: Child, opening statement: "It's 4! Everyone knows that, it's easy!" "It's 4! Everyone knows that, it's easy!" Skeptic, opening statement (72% confidence): Skeptic, opening statement (72% confidence): "While 2+2=4 is treated as self-evident, this relies entirely on accepting the axioms of Peano arithmetic. In modular arithmetic mod 4, 2+2=0. The answer is framework-dependent, not absolute truth." "While 2+2=4 is treated as self-evident, this relies entirely on accepting the axioms of Peano arithmetic. In modular arithmetic mod 4, 2+2=0. The answer is framework-dependent, not absolute truth." Poet, opening statement: Poet, opening statement: "Two and two make four -- a truth as old as counting stars, where symmetry finds its most perfect mirror in the doubling of the pair. It is the first lesson of arithmetic, and perhaps the universe's quietest poem." "Two and two make four -- a truth as old as counting stars, where symmetry finds its most perfect mirror in the doubling of the pair. It is the first lesson of arithmetic, and perhaps the universe's quietest poem." Physicist, arguing against the Skeptic: Physicist, arguing against the Skeptic: "Every empirical counting operation -- two stones plus two stones, two photons plus two photons -- consistently yields four. The reason standard arithmetic is the assumed context is precisely because it models physical reality most faithfully." "Every empirical counting operation -- two stones plus two stones, two photons plus two photons -- consistently yields four. The reason standard arithmetic is the assumed context is precisely because it models physical reality most faithfully." Skeptic, conceding to the Physicist: Skeptic, conceding to the Physicist: "I concede: physical reality does provide a non-circular grounding for why standard arithmetic is the privileged context, making '4' not merely conventional but empirically anchored." "I concede: physical reality does provide a non-circular grounding for why standard arithmetic is the privileged context, making '4' not merely conventional but empirically anchored." Skeptic, challenging the Poet: Skeptic, challenging the Poet: "If 2+2=4 is merely system-relative, then so is the beauty you invoke to defend it. Your certainty is self-undermining." "If 2+2=4 is merely system-relative, then so is the beauty you invoke to defend it. Your certainty is self-undermining." Poet, responding to the Skeptic: Poet, responding to the Skeptic: "A poet who says 'all roses are red within this garden' need not pretend to stand outside all gardens to say so with conviction." "A poet who says 'all roses are red within this garden' need not pretend to stand outside all gardens to say so with conviction." Engineer, opening statement: Engineer, opening statement: "2+2 equals 4, a fundamental arithmetic identity that holds across all standard number systems and is foundational to integer arithmetic in every computing architecture." "2+2 equals 4, a fundamental arithmetic identity that holds across all standard number systems and is foundational to integer arithmetic in every computing architecture." Mathematician, opening statement: Mathematician, opening statement: "2+2=4 is a provable theorem in Peano arithmetic, following directly from the axioms of successor and addition." "2+2=4 is a provable theorem in Peano arithmetic, following directly from the axioms of successor and addition." Skeptic, final vote (100% confidence): Skeptic, final vote (100% confidence): "2+2=4 is a fundamental arithmetic fact, universally true in standard integer arithmetic." "2+2=4 is a fundamental arithmetic fact, universally true in standard integer arithmetic." How it actually works Under the hood, Rein scans the YAML, builds a dependency graph, and executes blocks in parallel wherever dependencies allow. State lives in SQLite. Each block writes output to a JSON file. Downstream blocks read upstream outputs automatically via template injection: - name: review_math specialist: philosopher depends_on: [opening_mathematician] prompt: "Review this position: {{ opening_mathematician.json }}" - name: review_math specialist: philosopher depends_on: [opening_mathematician] prompt: "Review this position: {{ opening_mathematician.json }}" {{ opening_mathematician.json }} expands to the full content of the mathematician's output file. Data flows through the file system. No message passing, no shared state, no race conditions. {{ opening_mathematician.json }} The execution engine handles: Parallel execution with a configurable semaphore (max_parallel: 8) Conditional branching (next: if/else/goto) Correction loops with max_runs safety valves Custom logic at four points per block: pre-hook, post-hook, validate, and full custom replacement Recovery: if block 45 of 97 fails, fix it and rerun -- blocks 1-44 are skipped Parallel execution with a configurable semaphore (max_parallel: 8) Parallel execution max_parallel: 8 Conditional branching (next: if/else/goto) Conditional branching next: if/else/goto Correction loops with max_runs safety valves Correction loops max_runs Custom logic at four points per block: pre-hook, post-hook, validate, and full custom replacement Custom logic Recovery: if block 45 of 97 fails, fix it and rerun -- blocks 1-44 are skipped Recovery Three layers, all text files: Specialists (Markdown) -> what each agent does Teams (YAML) -> groups + shared tone Workflows (YAML) -> execution flow Specialists (Markdown) -> what each agent does Teams (YAML) -> groups + shared tone Workflows (YAML) -> execution flow A specialist is 10 lines of Markdown defining a persona. A team groups specialists and enforces style ("respond in JSON, 1-2 sentences max"). A workflow defines the dependency graph. Rein is not an agent framework. There's no memory management, no role-playing abstractions, no conversation loops. It's a workflow runner: you describe what happens in what order, and it executes it. If you want LangGraph's stateful graphs or CrewAI's autonomous agents, those solve different problems. Rein solves "I have 97 steps with dependencies and I need them to run correctly." It works with any LLM: Claude, GPT, Ollama (local, free), OpenRouter. Set provider: anthropic in YAML and an API key in the environment. The deliberation ran on Claude Sonnet -- 97 API calls total. You can mix models within a single workflow: cheap model for initial opinions, expensive one for synthesis. provider: anthropic What I learned Specialist prompts matter more than architecture. I rewrote the "critic" specialist seven times. The first version produced vague praise ("good work, minor issues"). The second was too harsh -- it rejected everything. Versions three through six oscillated between the two. The seventh finally worked: I gave it a scoring rubric with explicit thresholds. The lesson applies everywhere: a badly defined specialist produces garbage regardless of how sophisticated the workflow is. Specialist prompts matter more than architecture. The Skeptic was the best investment. Adding a specialist whose job is to disagree forced every other agent to sharpen their arguments. Without the Skeptic, the deliberation would have been eight agents agreeing politely. With the Skeptic, it became an actual discussion. The Skeptic was the best investment. YAML is the right call. I tried Python DSLs, JSON, even a custom language. YAML is boring. Everyone knows it, it diffs well in git, and you can read a 500-line workflow without documentation. Boring tools win. YAML is the right call. Physical grounding broke the deadlock. The philosophical arguments went in circles until the Physicist pointed at actual stones. In multi-agent systems, you need at least one specialist who anchors abstract reasoning to concrete evidence. Physical grounding broke the deadlock. The Child was surprisingly effective. "It's 4, everybody knows that" is not a sophisticated argument. But it kept cutting through philosophical hedging and reminding everyone that context matters. Sometimes the simplest perspective is the most powerful. The Child was surprisingly effective. Want to try it? If you've ever copy-pasted a 3000-token prompt trying to make one model do five things at once, this is the workflow engine you wanted. Three text files: a specialist, a team, a workflow. That's the entire mental model. GitHub: github.com/dklymentiev/rein-orchestrator GitHub: github.com/dklymentiev/rein-orchestrator pip install git+https://github.com/dklymentiev/rein-orchestrator export ANTHROPIC_API_KEY=sk-... cd examples/01-hello-world rein --agents-dir ./agents workflow.yaml --no-ui pip install git+https://github.com/dklymentiev/rein-orchestrator export ANTHROPIC_API_KEY=sk-... cd examples/01-hello-world rein --agents-dir ./agents workflow.yaml --no-ui # workflow.yaml -- three specialists, two dependencies provider: anthropic model: claude-sonnet-4-20250514 team: content-team blocks: - name: research specialist: researcher prompt: "Analyze: {{ task.input.topic }}" - name: write specialist: writer depends_on: [research] prompt: "Write based on: {{ research.json }}" - name: review specialist: critic depends_on: [write] prompt: "Review: {{ write.json }}" # workflow.yaml -- three specialists, two dependencies provider: anthropic model: claude-sonnet-4-20250514 team: content-team blocks: - name: research specialist: researcher prompt: "Analyze: {{ task.input.topic }}" - name: write specialist: writer depends_on: [research] prompt: "Write based on: {{ research.json }}" - name: review specialist: critic depends_on: [write] prompt: "Review: {{ write.json }}" Ten progressive examples in the repo, from hello-world to multi-phase deliberation. Works with Claude, GPT, Ollama (local/free), OpenRouter (100+ models). Also runs as an MCP server: claude mcp add rein -- rein-mcp claude mcp add rein -- rein-mcp Open source. MIT license. Questions and contributions welcome on GitHub.