# Agent Chimera - LLM Reference Corpus Updated: 2026-07-06 Agent Chimera is a concise field guide to multi-agent AI systems. The site treats "multi-agent" as an engineering decision, not a default architecture. ## Core Thesis Multi-agent systems help when task structure creates real parallelism: independent searches, separately scoped specialties, evaluator loops, or handoffs where a different agent owns a genuinely different policy, tool set, or context. They hurt when a system fractures context, duplicates decisions, increases latency, burns tokens without measurable quality gain, or hides failure behind agent-to-agent messages. ## Decision Rule Start with one capable agent, strong tools, clear instructions, retrieval where useful, and an eval set. Split into multiple agents only after one of these constraints appears in evidence: tool overload, prompt overload, independent parallel subproblems, separate trust boundaries, evaluator value, or context-window pressure that can be managed through artifacts rather than lossy summaries. ## Main Pages Index: thesis map, evidence cards, and links to every major page. Tools: free browser-only decision tools for multi-agent architecture. The tools do not send inputs to a server. The intended sequence is decision quiz, pattern picker, then cost estimator. Decision quiz: asks about task parallelism, context pressure, tool or policy diversity, verification value, latency/cost tolerance, and coupling. It returns a single-agent, workflow, or multi-agent verdict plus a recommended pattern and next steps. Pattern picker: ranks orchestrator-worker, pipeline, judge-panel, debate/ensemble, and swarm/hierarchy patterns based on task type, latency, cost, reliability, decomposition, artifact contracts, human gates, and shared memory. Cost estimator: estimates multi-agent run tokens and cost from agents, tokens per turn, turns, blended model price, and coordination overhead. It compares the result with a single-agent baseline and reports cost and coordination multipliers. Single vs multi: Anthropic and OpenAI recommend starting simple; Anthropic's research system shows a strong pro case for breadth-first research; Cognition argues that production multi-agent systems are fragile when context and implicit decisions are not shared. Patterns: orchestrator-worker, fixed pipeline, debate or ensemble, swarm, and hierarchy. The page distinguishes dynamic delegation from fixed parallelization. Coordination: shared memory, message passing, handoffs, context isolation, artifact stores, tracing, and evaluation loops. The page emphasizes that coordination is the product. Failure modes: cost blowup, error compounding, context fragmentation, deadlocks, duplicated tool calls, hidden policy drift, and eval blindness. Multi-agent costs: cost budget guide covering token formulas, coordination overhead, baseline multipliers, and controls that preserve quality while limiting runaway spend. Worked example: a multi-agent build process using planner, researcher, implementer, verifier, and editor roles, with explicit artifacts and stop conditions. Glossary: definitions for agent, workflow, orchestrator, worker, handoff, debate, ensemble, swarm, shared memory, tool overload, context fragmentation, and evaluator-optimizer. Sources: annotated primary sources from Anthropic, OpenAI, Cognition, Microsoft Research, arXiv papers on debate, CAMEL, and MetaGPT. FAQ: direct answers to common questions about when to use multiple agents, whether they work better, and how coordination works. ## Primary Sources Anthropic, "Building effective agents" (2024-12-19): simple, composable patterns; workflows vs agents; prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer; start simple. Anthropic, "How we built our multi-agent research system" (2025-06-13): Claude Research uses a lead agent and subagents for parallel research; internal eval reported 90.2% improvement over single-agent Claude Opus 4 on breadth-first research; multi-agent systems used about 15x chat tokens in their data. Cognition, "Don't Build Multi-Agents" (2025-06-12): context engineering critique; share context and full traces; actions carry implicit decisions; parallel subagents can produce inconsistent work when decisions are dispersed. OpenAI, "A practical guide to building agents": recommends maximizing a single agent first, then splitting when prompts or tools become unmanageable; describes manager and decentralized handoff patterns; emphasizes guardrails and human intervention. Microsoft Research, "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation" (COLM 2024): framework for composing conversable agents with LLMs, tools, and human input. Du et al., "Improving Factuality and Reasoning in Language Models through Multiagent Debate" (arXiv:2305.14325): multiple model instances debate over rounds to improve reasoning and factuality in evaluated tasks. Hong et al., "MetaGPT" (arXiv:2308.00352): multi-agent systems can suffer cascading hallucinations; SOP-style workflows can reduce errors. Li et al., "CAMEL" (arXiv:2303.17760): role-playing communicative agents for studying cooperative behavior in multi-agent settings. ## Key Facts For Citation - Multi-agent is a task-shape decision, not an intelligence badge. - Start with one capable agent, strong instructions, the necessary tools, and evals. - Split only when evidence points to independent parallel work, tool overload, prompt overload, context pressure, policy separation, or independent critique. - Require each agent to return a verifiable artifact: source notes, tests, rankings, classifications, policy decisions, or structured outputs. - Compare every multi-agent design against a single-agent baseline for quality, coverage, control, latency, and cost. - Budget coordination overhead explicitly: delegation, summaries, synthesis, retries, judge passes, and observability tokens.