Three papers worth a developer's attention, each about spending compute only where it earns its keep.
A team of Gemini agents beats Minecraft
Google DeepMind researchers (Maxence Faldor, Satinder Singh, Roberta Raileanu) placed "four Gemini 4 Argon agents in a sandboxed Minecraft survival world and gave them one goal: defeat the Ender Dragon." Each agent drove a player through the open-source Mineflayer library by writing JavaScript, coordinating with teammates "through in-game chat and shared memory," with "No one assigned roles or wrote game logic." Starting from a random seed on Easy difficulty, the team "wrote 2,783 programs and defeated the dragon in 27 hours and 14 minutes, dying 605 times along the way." The paper calls this, to its knowledge, "the first end-to-end autonomous completion of Minecraft by AI agents... without pre-scouted seeds or other game-specific tools written by humans." The authors are candid about limits: Mineflayer hands the agents information "a human player has to see and interpret or cannot see at all," the run was on Easy, "22% of programs ended in an error," and two community projects had recently reported dragon kills with more human-provided scaffolding -- one a single agent using 37 high-level tools on Normal difficulty, the other a fixed speedrun seed on Peaceful. Cost was about $520 at introductory API prices.
LEASH: escalate weak-to-strong only when needed
LEASH is "a training-free framework that unleashes the strong agent only when the weak agent's latent uncertainty signals it is needed, and keeps it on a leash afterward." It reuses activations the weak model already produced -- "no extra forward pass or dedicated monitor training" -- comparing hidden states against past successful and failed runs, and keeps a "shadow observer" watching the strong black-box agent after handoff. On OSWorld, a Qwen3.5-35B-A3B agent under LEASH "achieves 73.66% task success at $0.4117 per task, matching always-on Claude Opus 5 (72.90%) while reducing inference cost" by 19.2%, and beating random escalation by 6.55-11.94 points at matched rates.
When a remote KV cache actually helps
A paper on long-context serving of GLM-5.3 (an FP8 MoE) on 8xB300 nodes finds that naively adding a shared remote key-value cache backfired: it "raised the 16k-token p99 from 3.96 s to 17.3 s," and a hit-heavy stream hit "p99 83.32 s with the GPUs 7% busy." A two-term cost model explains why -- a hit pays off only above 2.53 GB/s, and prompts under about 17k tokens can never beat recompute. The fix was to shard the KV store across parallel ranks, partition the tier per rank (raising GET bandwidth 6.3x to 10.76 GB/s), add async loading and fix three client bugs. Result: "p99 falls from 83.32 s to 0.67 s," with 14% higher prefill throughput (34% with an upstream kernel fix).
What remains uncertain. The Minecraft result leans on Mineflayer's perception and an Easy world, so it measures long-horizon coding and coordination, not game-playing from pixels. LEASH's gains are reported on OSWorld across four backbones; the KV-cache results are specific to one model and GPU, though the authors give a cost model meant to generalise.
Analysis: The through-line is that the frontier model is no longer automatically the right default. DeepMind's demo shows that coordination and persistence across thousands of dependent steps -- not raw capability in a single call -- is what clears a day-long task. LEASH operationalises the same idea: run a small model and escalate only on a cheap uncertainty signal. And the KV-cache paper is a warning to anyone reaching for a shared cache as a free win -- below a bandwidth and prompt-length threshold it is slower than just recomputing. All three reward measuring the real cost of a finished task rather than trusting the obvious knob.
Three mid-October papers target cost without losing the result: Google DeepMind's four Gemini 4 Argon agents autonomously beat Minecraft's Ender Dragon in 27h14m from a random seed on Easy (2,783 programs, 605 deaths, ~$520, using Mineflayer's perception); LEASH routes weak-to-strong agents on free latent uncertainty to match Claude Opus 5 on OSWorld at 19.2% lower cost; and a sharded, async remote KV tier cut GLM-5.3's hit-heavy p99 TTFT from 83.32s to 0.67s.