Agency by Mozilla vs FlashInfer 2.0

“The primitive here is clean: a browser-native agent runtime that binds to Firefox's internals and exposes MCP-compatible tool interfaces, all local. No cloud hop, no screenshotting your desktop and sending it to Anthropic. The DX bet Mozilla made is right — run in-process in the browser where DOM access is first-class, not bolted on from outside. The moment of truth is whether the MCP tool registration is actually ergonomic or if it buries you in schema boilerplate, and the repo suggests the latter needs polish. Still, this is a real primitive, not a wrapper — Mozilla is giving developers a composable base that a Playwright-over-CDP weekend project genuinely cannot replicate, because the privacy guarantees come from architecture, not policy.”

“The primitive here is a CUDA kernel library for attention computation and KV-cache management — not a platform, not a wrapper, an actual low-level building block you can drop into vLLM or SGLang. The DX bet is correctness and composability over abstraction: they expose the knobs (speculative decoding thresholds, LoRA batching configs) without hiding them behind a config YAML that pretends the complexity doesn't exist. The moment of truth is swapping in the FlashInfer attention backend in an existing serving stack, and from what the repo shows, that's genuinely a few lines. The 40% latency claim needs a methodology cite — they show specific token generation benchmarks on H100s with prefill/decode separation, which is at least a real number attached to a real setup, not a vibe. This is infrastructure that a competent team could not replicate in a weekend; the CUDA work is deep and the speculative decoding integration is non-trivial. Ships because the craft is demonstrably in the kernels, not the landing page.”

“Category is browser automation agents; direct competitors are Anthropic Computer Use, OpenAI Operator, and Playwright-based agent wrappers. The scenario where this breaks is any user who needs a capable frontier model baked in — Agency gives you the runtime plumbing but you still have to bring your own model, and local models are still embarrassingly bad at browser task reasoning compared to GPT-4o. What kills the cloud alternatives here is regulatory pressure on enterprise data handling, which is real and accelerating — that's the thesis that survives. Mozilla ships this, it gets traction in privacy-sensitive enterprise and research contexts, and the cloud agents find their growth capped in regulated industries. I'd call this a genuine ship for the niche it's targeting, not a universal recommendation.”

“Category is LLM inference optimization, direct competitors are FlashAttention-3, vLLM's built-in attention kernels, and NVIDIA's TensorRT-LLM — none of which are sleeping. The 40% latency claim is real in a narrow regime: it applies to specific decode-heavy workloads on Hopper-generation GPUs with prefill-decode disaggregation; swap in an A100 cluster doing long-context prefill and the number shrinks. What kills this in 12 months is not a competitor — it's NVIDIA shipping optimized kernels directly into cuDNN or the next-generation attention primitives landing in TensorRT-LLM, at which point the delta collapses. What earns the ship anyway: multi-LoRA batching at scale is a genuinely underserved problem that the big players haven't prioritized, and Together AI has production traffic to validate these claims against real workloads, not synthetic benchmarks. The open-source release is credible signal that they're playing for ecosystem, not just headlines.”

“The falsifiable thesis here is: within 3 years, regulatory and user-trust pressure will make cloud-routed browser agents legally or commercially unacceptable in enough markets that local-first agent runtimes become the default for sensitive workflows — healthcare, legal, finance, government. Agency is early to that specific bet, and being a Mozilla project means it rides the browser-vendor trust signal that no startup can buy. The second-order effect nobody's talking about: if Agency becomes the standard runtime for Firefox-native agents, Mozilla gets to define what MCP tool permissions look like in a browser context, shifting standards power back toward an open-standards body and away from the model providers. The dependency that has to hold is that local model capability closes the gap with cloud fast enough — Gemma 3 and Qwen3 suggest it's on track.”

“The thesis here is specific and falsifiable: inference compute will remain the dominant cost in LLM deployment for at least the next three years, and kernel-level optimization will continue to yield meaningful gains even as hardware scales. What has to go right is that the prefill-decode disaggregation architecture becomes the dominant serving pattern — if monolithic batching stays standard, FlashInfer's architectural assumptions become a liability rather than an asset. The second-order effect that matters most isn't latency reduction for Together AI's own platform — it's that cheap, reliable multi-LoRA serving changes the economics of fine-tuning. If you can serve 50 LoRA adapters off one base model at acceptable latency, the cost of domain-specific fine-tuning drops by an order of magnitude, which shifts power toward the fine-tuning layer and away from base model providers. FlashInfer is riding the prefill-decode disaggregation trend, and it's on-time rather than early — vLLM and SGLang have already moved this direction, which means the ecosystem is ready to absorb this rather than resist it.”

“There is no buyer here, which is the whole problem — Mozilla is a nonprofit shipping open-source infrastructure, not a business, and that's fine for what it is, but framing this as a product review misses the point and also confirms the skip. Any startup trying to build on top of Agency inherits Firefox dependency, local model constraints, and a framework maintained by a nonprofit with a historically mixed record of developer-facing project continuity (see: Firefox OS, Servo, Pocket). The moat question answers itself: Mozilla can't own a market position because they're not trying to, and any company that builds a product layer on this is one browser vendor decision away from a breaking change. If you're a developer building privacy-first browser tooling, this is interesting infrastructure. If you're trying to build a business on it, that's the skip.”

“The buyer here is infrastructure engineers at companies running self-hosted LLM inference at scale — a real buyer with a real budget (GPU compute costs), not a vague enterprise persona. The open-source release is a distribution play, not a charity: Together AI captures value through their managed inference platform, where FlashInfer improvements directly reduce their per-token compute cost and become a credible differentiator in a market where Fireworks, Groq, and Anyscale compete on latency benchmarks. The moat question is the hard one — open-sourcing the kernel library means competitors can adopt it too, so the defensibility is execution velocity and production integration depth, not the code itself. What happens when NVIDIA ships this natively is the real stress test, and the honest answer is that Together AI's moat shifts entirely to their managed platform and the workflow integrations built on top of it. Still a ship because the business logic is coherent: they're using open source to build pipeline credibility while monetizing on the managed layer, which is a proven playbook.”