A **Block Engineering** benchmark post from **August 6, 2026**, signed by **Atish Patel**, about **Buzz** — the human + agent workspace launched on July 21 — asking a cost question: which agent team is **the cheapest one that reliably succeeds**? Three findings. **(A) A negative result, published in full**: on **Terminal-Bench 2.1**, **twelve team compositions** (pairs, triads, cheap swarms under a *frontier* model) were pitted against the solo agent each was built around, and **none beat it at equal cost**. The explanation is structural — a task that finishes in minutes *"doesn't have enough structure to divide"*, and *"More agents mostly buys you the cost of explaining it twice"*. **(B) The horizon reverses the result**: on **Long-Horizon Terminal-Bench** (44 tasks, one task worth hours of work, same lead **GPT-5.6 Sol** at *high* effort), solo finishes 15 tasks for 59.1%, +2 QuickBees 19 for 64.1%, +1 QuickBee +1 WorkerBee 19 for 69.5%, **+2 WorkerBees 20 for 71.5%** — a **+12.4-point** gain, of which 11.4 comes from tasks carried to completion. *"Same seats, opposite result, because the work is a different shape."* These runs ran at **3× the timeout**, solo included. **(C) Beyond a threshold, price stops buying quality**: solo on Terminal-Bench 2.1, **Opus 5 at *xhigh* effort is the most expensive run ($140.63) for 75.0%**, trailing six runs ranging from $20.08 to $109.82 and 79.5% to 88.4% — the stated cause is over-reasoning that drove 17 of 88 tasks to timeout. Among the six best runs, **a 5.5× price gap for an 8.9-point score gap**: *"choosing between them is not a quality decision at all. It is a budget decision."* The post proposes a taxonomy it owns as *ad hoc* — **QuickBee**, **WorkerBee**, **SmartBee**, plus the human as *"honorary bee"* — and two team forms, the permanent **Hive** that remembers your preferences and the disposable **Swarm** that remembers the project. Conditions: everything runs on **Harbor**, against real Buzz agents on a **live** relay, **one attempt per task, no retry**, prices fixed as of **2026-07-30**.
#Buzz#Block#agent teams
- **Atish Patel** — *« Building AI solutions @ Block »* · auteur unique du billet · publié le **6 août 2026** sur `engineering.block.xyz`.
Announcement from **Meta AI Research** published on **August 5, 2026** (stated reading time: 4 minutes, no individual byline): **Muse Code** in beta, *« a terminal coding agent »*, and the model that powers it, **Muse Spark 1.2**. Meta itself frames the launch: *« This marks our next step toward the frontier, with larger and much more capable models on the way. »* **Three architectural elements on the harness side.** **Asynchronous background agents** that *« remain active throughout each session, rather than being spawned for individual tasks »*, avoiding redundant information gathering and reducing the need for steering. A **local event log** where *« every model call, tool run, approval, and edit is appended »*, making the runtime a system that is *« replay-exact and restart-safe »*, able to resume exactly where it left off after a crash. And **three skills shipped out of the box**: `/plan` (turns a task into a plan submitted for approval), **`/grill`** (stress-tests the plan *« until it holds up »*), and `/goal`. **On the model side**, Meta claims **model-harness co-training** (*« to maximize harness compatibility »*, with harness trajectories sampled via rejection sampling and recipe optimizations for goals, compaction, and sub-agents), **long-horizon** training (whole-repo generation, end-to-end projects, self-research, with planning, goal conditioning, and context compaction), and a **self-improvement loop** where Muse Spark 1.1 generates the environments and instruction templates and then grades candidate solutions, producing a training set for the 1.2. **What the published charts show**, without the text commenting on it: the four comparisons — Terminal-Bench 2.1, DeepSWE 1.1, an internal Meta benchmark, and the GPU kernel optimization case study — place **Muse Spark 1.2 behind Opus 5 in all four cases**, including on Meta's own proprietary benchmark (70.6% versus 79.4%) and on the case study, where the model finishes fourth out of six (+68.7% versus +74.0%). **A reading caution on the version gain**: on the two public benchmarks, 1.1 is measured with `mini-swe-agent` and 1.2 with Muse Code, so the 6.7-point gap conflates model and harness. On the internal benchmark, the only comparison where no harness is mentioned, the 1.1 → 1.2 gap drops to **2.3 points**.
#Meta AI Research#Muse Code#Muse Spark 1.2
**Meta AI Research** — publication institutionnelle sans auteur nommé · sur `research.meta.ai`. Le billet renvoie à un **rapport** pour la méthodologie d'évaluation · non repris ici.