Two threads on Hacker News tell the same story from different angles. The first, with 204 points, documents how GPT-5.6 Sol closed a problem open for roughly 30 years in convex optimization, using a prompt and parallel sub-agents. Six days earlier, on July 12, we covered the proof of the Cycle Double Cover Conjecture: same model, same pattern.

The second thread, with 121 points, is Charles Azam’s test. He put Fable 5 and GPT-5.6 Sol on the same NP-hard problem (designing a fiber optic network with an estimated search space around 10^1223 solutions) with 30 minutes of budget each. The result is instructive for anyone orchestrating agents.

Why it matters to you. Fable 5 beat Sol in standard configuration (average 32.386 vs 34.261) with much tighter variance: 319 points of swing vs 1.958. The stronger model on paper lost. Azam also shows that the /goal command, present in Claude Code and Codex, restructures the control loop instead of intensifying it. In Claude Code it delegates to a separate evaluator model that reads the conversation and says yes or no. In Codex it becomes persistent state in SQLite, with dedicated tools. Two different architectures under the same command.

Orchestration beats the model. If you’re delegating hard tasks to an agent, the control loop is where you win or lose.

In detail

The Reddit thread doesn’t give the full proof text, but the title and discussion on Hacker News (204 points, 106 comments) indicate that GPT-5.6 Sol produced a closed proof for a convex optimization problem open for roughly 30 years. The structure mirrors the July 12 result on the Cycle Double Cover Conjecture: the model orchestrates sub-agents in parallel, each with a problem fragment, and combines results into a proof verified by a mathematician.

Azam’s test gets practical. The KIRO problem is real: design a fiber optic network for Grenoble, Nice and Paris, where each tower appears once, loops must stay under 30 terminals, and the goal is to minimize cable length. Azam solved it in C++ in 2018, so he has a human baseline. The search space, even in a simplified version (19 loops from 28 terminals, no branches), is on the order of 10^1223.

The numbers say two things. First, Fable 5 was more consistent: in three runs without /goal, results fall in a 319-point range (32.197 to 32.516). Sol in the same exercise spans 1.958 points (33.581 to 35.539). Second, /goal won 4 of 6 runs but worsened the average for both models. Sometimes it finds a better basin, sometimes it gives a wrong idea time to mature.

The most useful detail is under the hood. /goal in Claude Code is a stop hook at session level: after each turn of the main model, a small model (Haiku) reads the conversation and decides if the objective is met. It can’t use tools or inspect files (it judges only from the transcript). In Codex (version 0.144.4, open source), /goal is persistent: the TUI saves the objective to SQLite with state and budget, and the model receives dedicated tools (create_goal, get_goal, update_goal). When the thread goes idle with an active goal, Codex injects a continuation turn with a completion audit. Two philosophies: Claude delegates to another model, Codex puts state in the system.

Limits are stated. Azam tested one problem, with six runs each for the two top models and 30 minutes of arbitrary budget. Claude Code’s implementation is closed source, so it relies on what Anthropic documents. For those building agents, the practical consequence is clear: before hunting for the strongest model, evaluate how your control loop handles the objective. The course on orchestrating multiple agents and the playbook for delegating a multi-step task build exactly this competence.