Grok 4.20: 4-Agent Multi-Agent System (announcement)

TL;DR

A pointer-style tweet broadcasting xAI’s Grok 4.20 launch (Feb 17, 2026), which ships a native “4 Agents” multi-agent architecture as a consumer product: four specialized replicas of the same underlying model — Grok (captain/synthesizer), Harper (research/fact-checking via X firehose + web), Benjamin (math/code/logic verifier), Lucas (contrarian/creative) — run in parallel, debate each other in structured rounds, and only emit a final answer after consensus. xAI reports a 65% hallucination reduction (~12% → ~4.2%) and claims marginal cost of 1.5–2.5× a single forward pass (not 4×) by sharing weights, KV-prefix, and input context across agents on Colossus. Relevant to the team as a production-scale datapoint for inference-time scaling via parallel specialized agents — the same axis the wiki has been tracking through K2.5’s PARL (learned orchestrator), MiroThinker (interaction-depth), and RLM (programmatic scaffold). Caveat: the tweet itself is an announcement, not a paper; what’s reported below is what xAI / secondary write-ups claim — no public technical report exists.

Key claims

Grok 4.20 ships four specialized agents (Grok, Harper, Benjamin, Lucas) that run in parallel on every sufficiently complex query, debate, and only synthesize after internal peer review [tweet OP].
One agent (Lucas) is explicitly trained to be a contrarian / disagree with the other three [tweet OP, secondary write-ups].
xAI reports a 65% drop in hallucinations (~12% → ~4.2%) attributed to the peer-review mechanism between agents [secondary write-ups citing xAI].
The four agents share model weights, prefix/KV cache, and input context, giving marginal cost ~1.5–2.5× a single pass rather than 4× [secondary write-ups; xAI Colossus claim].
Heavy-tier users can scale the same architecture to 16 agents on a single prompt [secondary write-ups].
xAI exposes the system via grok-4.20-multi-agent with an agent_count parameter on the API [xAI docs].

Method

The tweet is a launch announcement and contains no technical detail beyond “4 agents, think → debate → synthesize”. Reconstructing from secondary sources (xAI docs + write-ups), the inference-time workflow is: (1) Task decomposition by the Captain (Grok) into sub-tasks routed to specialists; (2) Parallel independent thinking — all four receive full context plus their specialized lens and generate analyses concurrently; (3) Multi-round structured debate — Harper grounds factual claims in real-time X/web data, Benjamin checks logical/numerical consistency, Lucas surfaces missing perspectives and biases, with iterative cross-correction until consensus or flagged uncertainty; (4) Synthesis — Captain aggregates and emits one user-visible answer. Key reported optimization is that the four agents are not separate models but four personas instantiated on the same underlying ~3T-parameter MoE weights with shared prefix cache, plus structured/RL-trained-short debate rounds (xAI claims pre-training-scale RL is used for orchestration efficiency).

Results

Hallucination rate: ~12% → ~4.2% (65% relative reduction), per xAI’s reported numbers.
Arena ELO: xAI claims 1505–1535, in Gemini 3 Pro territory and above Claude Opus 4.5 / GPT-5 (secondary write-ups; not independently verified).
ForecastBench: reported #2 globally, above GPT-5 / Gemini 3 Pro / Claude Opus 4.5 (xAI claim).
Alpha Arena (stock-trading sim): +34.59% return while competitors posted losses (xAI claim).
API: grok-4.20-multi-agent with selectable agent_count (default 4, up to 16 for Heavy).

All numbers are vendor-reported and not from a peer-reviewed source.

Why it’s interesting

Sits as a production-scale datapoint on the parallel-agent axis of [[inference-time-scaling]] — distinct from RLM (untrained programmatic scaffold), MiroThinker (sequential interaction-depth), and K2.5’s PARL (learned orchestrator policy over ≤100 sub-agents). Two specific things worth tracking: (1) the weight-shared, cache-shared implementation is what makes the 1.5–2.5× marginal cost claim plausible — if true, this is a strong argument that “personas on shared weights with shared prefix” is a much cheaper deployment unit than “N independent model replicas”, which directly applies to any agent-swarm setup the team might build; (2) the named-failure-modes pattern (Lucas-as-trained-contrarian) is a different solution to the “serial collapse / spurious parallelism” failure modes that K2.5’s PARL addresses with shaped rewards — xAI uses role specialization, K2.5 uses reward shaping. No public ablation distinguishes which works better.