This post is a Feynman-style explanation, originally written in Chinese, of Dahal et al., "Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning" (arXiv:2609.38147, 2026-09-29), by researchers from Meta Superintelligence Labs, FAIR, and Princeton.
The core analogy
The author opens with a home-renovation story: eight skilled workers stalled for a week because every small decision required asking the foreman, who was shouted at two hundred times a day. The fix was a ten-minute daily stand-up meeting. The paper's claim maps onto this: when agents handle long tasks, the bottleneck is not the "working hand" but the chaotic, ad-hoc way "what to do next" decisions are made. The paper calls these control decisions—which draft to continue from, when to backtrack, when to stop. The mainstream approach, direct control, stuffs the entire history into the model at every step; as tasks grow, this floods the model's attention and produces inconsistent, unrecoverable decisions.
Tower and wrench: the two-role architecture
The paper splits an agent into:
- Workers: execute concrete task-level work (writing code, reasoning on assigned steps).
- Control tower: never touches the wrench. Its job has four parts, like a tower operator's day: 1. Assess: what facts has this run established so far? 2. Propose: what are the available next options—refine the current draft, switch routes, or revive a failed branch? 3. Evaluate: weigh each option against the remaining budget (time, compute, tokens). This is what separates "management" from "advice." 4. Dispatch: assign the chosen task to a worker with a clear work order pinned to established facts.
- Attention dilution: the critical failure buried 20,000 words deep gets averaged away in full replay.
- Cost: full-history replay makes the token bill balloon; the budget goes to recalling, not working.
- ProgramBench (long-horizon program refactoring, dense in control decisions):
- GPT-5.5 + meta-reasoning: 71.5% vs. production-grade Codex agent: 58.0% (+13.5 points).
- Opus 4.8 + meta-reasoning: 67.2% vs. production-grade Claude Code: 65.5%.
- Across abstract reasoning, multi-domain long-horizon reasoning, and proof-generation benchmarks: average gains of 3.6–4.2 points across three frontier models versus direct control.
- Baselines include production agents, research harnesses, and a purpose-built direct-control agent with the same workers and compute budget—isolating the variable of coordination. Notably, the paper honestly reports conditions where the method underperforms.
- Good news: with ample budget, meta-reasoning keeps improving while direct control plateaus—the tower converts added resources into more high-quality decision rounds because it never carries full history.
- Bad news: at small budgets, the tower's per-cycle cost (reading, proposing, evaluating all cost tokens) dominates—like holding three review meetings for a three-line message. Practical advice: meta-reasoning should not be a default; check task size and budget first. Small jobs, work directly; big jobs, use the protocol.
- Dahal et al. "Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning", arXiv:2609.38147, 2026-09-29
This is meta-reasoning: reasoning about how to organize the next round of reasoning, before the workers think—just as "metadata" is data about data. The tower and workers are decoupled, so either side can improve independently, mirroring coach/team, surgeon/scrub-nurse, and prefrontal-cortex/motor-cortex divisions of labor.
Carrying only the meeting minutes
Instead of replaying the full run history at every decision (the boss forced to watch 200 hours of meeting recordings), the controller carries a compact run account—conclusions, not process: what's done, what was rejected, which branch we're on. This addresses:
The author likens this to a ship captain's logbook—conclusions and rationale, not the voyage itself—and to human working memory. The account is maintained continuously by the Assess step; minutes culture is the product of retrospective culture.
The four-step cycle
Assess → Propose → Evaluate → Dispatch, looping until the task is done. The post notes how each step inverts the "let the model freestyle" intuition: Propose forces listing multiple options (turning gut decisions into multiple-choice); Evaluate makes budget a decision *input* rather than an afterthought; Dispatch delivers a concrete work order with a pinned starting point, not vague encouragement.
The numbers
When money is scarce, overhead is a burden
Three secrets in the artifact graph
The paper builds a graph of all artifacts (code segments, drafts, branches—who derived from whom, what was adopted):
1. Tower-run agents reuse early work more—assessments keep early artifacts alive in the account rather than buried in history. 2. Higher coverage of correct solutions—forced multi-option proposals spread the search. 3. Final gains are uneven—even in tower runs, some adopted choices pay off richly, others don't; management raises the floor and the average, not every single bet.
Together: the scores show *that* it works, the scaling curves show *when*, and the artifact graph shows *why*.
Closing: the first of three gates
The post frames coordination as only the first gate for strong long-task agents. The second is honesty (sticking to the plan during execution); the third is evolution (self-improvement without teacher scores or reward signals). What impressed the author most: everyday human common sense—stand-ups, minutes, triage, logbooks—translated seriously into a system and validated experimentally. Once common sense survives such testing, it stops being common sense and becomes a principle.