Korg2.4
Korg2.4 started as a reverse-engineering exercise, built in collaboration with Krish Chhajer. Grok 4.2 looked like it was running a native multi-agent collaboration system on every sufficiently complex query: not a user-facing orchestration framework like AutoGen or Swarm, but something baked into inference itself. Four specialized replicas of the same underlying model, collaborating in real time.
I wanted to know how much of that behavior you could recover without access to the weights, the routing code, or the training recipe. The working thesis is simple and uncomfortable: you do not always need a real multi-agent system. A lot of the upside may come from structured coordination, role separation, and handoffs. That pattern might even be learnable inside one model if you know what to target. That last part is the point. Architecture 3, internalized unified deliberation, was always the actual goal: one model that simulates the committee without five separate instances at inference time.
This is an architecture paper, not a post-training paper. The question is explicit coordination vs. learned internal specialization.
What Grok 4.2 looked like from the outside
From public signals in late March, the stack looked like four agents forming a production collaboration layer on hard queries: Captain, Harper, Benjamin, and Lucas. Captain decomposes the prompt, routes subtasks, resolves conflict, and ships the final answer. Harper handles search, evidence, and fact verification, with heavy use of live X and web data. Benjamin stress-tests math, code, and logic. Lucas adds divergent hypotheses, balance, and polish.
The hypothetical workflow is parallel analysis first, then internal peer review, then synthesis. Harper grounds claims. Lucas flags rigidity. Benjamin breaks fragile reasoning chains. Captain picks the winning arguments and writes one coherent output. OpenAI's prior work leans heavily on chain-of-thought and Codex-style parallel agents. xAI's apparent edge is speed and scale at inference time, plus a lot of RL on specialized roles. I cannot verify any of that from the inside. What I can do is build explicit baselines and measure whether committee-style coordination buys accuracy at a predictable cost.
Architecture 1
Iterations run on Qwen3-8B, with extrapolation planned to Qwen3-30B. Architecture 2 tests shared-weight conditional roles. Architecture 3 is where this was supposed to end. This page covers Architecture 1 only: the explicit-coordination baseline you can actually run without post-training budget.
Architecture 1 implements a fixed, externally orchestrated committee. Four role-specialized instances of Qwen3-8B, all in non-thinking mode, run in a deterministic five-stage pipeline: coordinator plan, researcher, analyst, critic, coordinator synthesis.
The coordinator converts the raw task into a structured plan at the start: restated problem, success criteria, risks, and a research brief. At the end it receives the research handoff, analyst draft, and critic feedback, then decides what evidence survives and how to synthesize the final response.
The researcher acquires and organizes evidence from the task context, passages, answer choices, and metadata. In tool-enabled settings it can use retrieval or search. The other roles only see its structured handoff.
The analyst produces the first complete candidate answer from the plan and research handoff. The draft is provisional and goes to the critic, not straight to the user.
The critic reviews the draft against the evidence and success criteria. It flags unsupported claims, missing steps, weak assumptions, and format issues. It does not replace the analyst. It applies adversarial pressure before the coordinator ships.
Specialization is fully externalized in inference-time topology. Every role is a distinct model instance with its own system prompt, reasoning trajectory, and communication boundary. That gives the clearest role separation and the highest inference cost: repeated prefills, repeated context transmission, redundant decoding across agents.
Results
HotpotQA · 75 examples · Qwen3-8B-AWQ
| System | Done | EM | F1 | sec/task | Calls | Tokens |
|---|---|---|---|---|---|---|
| Arch 1 committee (non-thinking) | 73/75 | 38.36% | 57.74% | 150.04 | 365 | 1.54M |
| Single-shot | 75/75 | 0.00% | 1.62% | 23.06 | 75 | 175K |
Head-to-head F1: committee wins on 68 tasks, single-shot on 4, tie on 1. The committee used roughly 8.8× the tokens, 6.5× the latency, and 5× the model calls. The gain is real. So is the bill.
GPQA Diamond · 198 examples · Qwen3-8B
| System | Done | Acc. | Full-set | sec/task | Calls | Tokens |
|---|---|---|---|---|---|---|
| Arch 1 committee (non-thinking) | 163/198 | 51.53% | 42.42% | 383.16 | 815 | 3.36M |
| Qwen3-8B non-thinking ref. | N/A | 39.30% | N/A | N/A | N/A | N/A |
| Qwen3-8B thinking ref. | N/A | 62.00% | N/A | N/A | N/A | N/A |
On the harder benchmark, the non-thinking committee clears the non-thinking reference but sits below the thinking reference. Same tradeoff: more accuracy on structured reasoning, paid for in calls, tokens, and wall clock.
Korg2.4 is not trying to clone Grok. It is trying to ablate what black-box multi-agent behavior actually requires. Architecture 1 proves that explicit coordination can move a small model far on multi-hop QA, even when single-shot collapses. Architecture 2 might recover some of that gain at lower inference cost. Architecture 3 is the destination: bake the committee into the weights so you stop paying for five passes on every query.
I did not get there. Architecture 3 needs post-training, not just prompting, and an honest run at that looks like roughly $20K in compute I do not have sitting around. So this project stops at the inference-time baseline: measure what explicit coordination buys you, document the cost, and leave a clean target for the version that actually internalizes it.