Fuse the bottleneck, price your agent platform, and verify what code ran
Two AI picks: Meta's Andromeda rewrite, where fusing kernels and co-locating index matching cut network traffic about 100x, and GitHub Universe, where agent platforms now come with multi-model routing and usage-based credits. One engineering pick: attestation-backed verification of running code on AWS Nitro Enclaves.
AI & Agent Engineering
ML systemsvendor-reported
Meta Andromeda: fusing model and system work into one GPU kernel cut network traffic about 100x
Meta rebuilt its ads retrieval engine around a mixture-of-experts model with about 300x fewer inference FLOPs than the 2024 version. Scoring, index matching and top-K selection now run in one fused GPU pipeline on a single server, which Meta says reduces network traffic by about 100x and processes over one billion ads per second.
Why it matters
At this throughput the bottleneck moved from arithmetic to memory bandwidth and then to the network, so each fix exposed the next limit. These are Meta-reported numbers on custom hardware, but the method of finding the real bottleneck first carries over.
Do this week
Before scaling a hot inference or scoring path, profile where bytes move, not only where compute goes. List the intermediate results that cross a memory or network boundary, and try co-locating the stage that produces them with the stage that consumes them.
GitHub Universe: Copilot app, multi-model Fusion and credit-based pricing for agent work
GitHub announced the Copilot app, Copilot projects for parallel multi-repo agent sessions, and Copilot Fusion, which chooses one or more models per task. GitHub says Fusion Max scored within two points of Opus 5.5 Max on the Artificial Analysis Coding Agent Index at half the cost. Enterprise customers can use all Copilot features without a separate subscription and pay for AI credits used.
Why it matters
Pricing is moving from per-seat to per-use, which makes agent cost something a team has to budget and monitor. The benchmark result is vendor-reported and has not been independently confirmed here.
Do this week
Before moving to credit-based billing, pull a month of agent usage per team and estimate cost per merged pull request. Set a per-team credit budget and an alert so a runaway agent loop shows up in days, not at invoice time.
AWS: proving which code ran with Nitro Enclave attestation and reproducible builds
Anchorage Digital and Turnkey describe running an open source transaction parser inside AWS Nitro Enclaves. The enclave signs responses, and clients compare attestation PCR measurements against values from reproducible builds. Anchorage open sourced awsnitroverifier, a Go library that separates malformed input from validation failures.
Why it matters
Network perimeters and contracts do not prove which binary handled a request. Reproducible builds make the expected measurements stable across rebuilds, so verification policy does not need constant updates. The post is a customer case on a crypto workload.
Do this week
Check whether your builds are byte-identical when run twice on different machines. If they are not, fix that first, because attestation or signature checks of any kind depend on a stable artifact hash.
Pick one service. Build it twice on two different machines or CI runners from the same commit, then compare the SHA-256 of the output artifact or container image.
Why. Identical hashes are the precondition for signing, attestation and cache reuse. Differences point to timestamps, paths or unpinned dependencies.
Measure. Record whether the hashes match and list each source of difference you find. Aim to remove the top one this week.
Classic / essential read
Continuous Integration
Martin Fowler · 30 min
It sets out the practices that keep a mainline releasable: a self-testing build, daily pushes to mainline and fixing broken builds immediately. Those matter more when agents multiply the number of changes.
One lesson. Measure how long your mainline stays red and how old your oldest unmerged branch is. Both numbers tell you how much integration risk you carry.
Find where bytes and money actually go before optimising. Profile data movement in hot paths, estimate agent cost per merged change, and confirm your builds are reproducible.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.