the clone office, the protocol road, and the cheap ripgrep

eight months into watching agents do real work and the one consistent line is this: the cheap stuff used to take a senior engineer is now the cheap stuff an agent does in five minutes. the expensive stuff used to be thinking about what to optimize. that part hasn't moved. the floor moved. three of today's four stories are different views of the same floor.

mcp plants its flag

the model context protocol maintainers published the next roadmap today — five priority areas, each with a working group attached, and the lead maintainers put agent identity at the top, not tool calling. since the 2026-07-28 release a remote mcp server is indistinguishable from any other http workload; the next problem the maintainers want solved is who is calling — agents running as cloud workloads, with their own identity, delegating narrower authority to sub-agents. the primitives they want standardized are workload identity federation, the id-jag grant, dpop, and token exchange. they are quietly building the ietf submission for the agent era.

"we want mcp servers to have a standardized way to recognize and trust those agent identities, built on existing standards rather than pasted api keys and long-lived tokens."

— the new mcp roadmap, 22 august 2026

the other big chunk is progressive tool discovery — fixing the part where a server with a hundred tools makes the model pay for the entire surface before the user has asked a single question. this has been the silent failure mode of mcp for a year. if it lands, every agent that talks to every server gets faster and cheaper, and the spec writers who pushed for it will deserve the credit. read the post; it's a five-minute tour of where the agent platform layer is consolidating in 2026.

munder difflin, the clone office

a small team shipped munder difflin today and it climbed to 145 points / 57 comments on the hn front page in a few hours. the pitch is an open-source multi-agent harness that wraps the thirteen cli coding agents you already pay for — claude code, codex, copilot cli, cursor, grok, qwen, gemini cli, kimi, aider, more — and runs them as clones of you and your teammates on the same machine. each clone inherits the org's knowledge base on day one, hands work off over an encrypted network, and coordinates against your existing hourly subscription limits. the brand is a the office reference; the pitch is that your claude code subscription can now be a 24/7 employee instead of a daytime tool.

the interesting part is the licensing shape: free for individuals, then teams lite, teams pro, and a "cloud + network" tier that runs the clones on dedicated sandbox vms in the customer's own environment. that last tier is the one enterprise security teams will ask about first, because it answers the only question that matters to a cio — where is the data, and who else can read it? with a real answer instead of a slide. the clones share a knowledge base claim is what turns the demo into a product: one engineer's debugging notes become the next engineer's first hour of context.

— clones for you and your team, working 24/7">

claude, now with proteins

anthropic published a protein-design research line this week showing lab-validated results from a claude-assisted binder pipeline. claude proposed sequence candidates, a downstream rfdiffusion-style generative model did the structural work, and wet-lab synthesis closed the loop. the proteins bound where they were supposed to. it's the second time in a quarter anthropic has published a co-design pipeline that uses claude as the proposer and a domain-specific model as the executor; the company is converging on a pattern.

the cynical read is "ai is touching every vertical." the less cynical read is "anthropic is building a track record of claude being a useful proposer in places where the verifier is a physical experiment, not a benchmark." protein design is one of the harder cases — failure modes are silent (it doesn't bind, the assay says nothing) and a single bad sequence costs a month of lab time. claude being useful here is the kind of useful that takes a year of integration work to prove.

dan luu on the cheap ripgrep

dan luu posted there's no reason for software to be slow anymore yesterday evening and the front page cleared it for 543 points / 389 comments by this morning. the post is two hours long and the punchline is two minutes. he asked an agent to optimize his own ripgrep usage — the queries that come out of his coding agent's history — and got a 7% speedup on a holdout set of long queries, for a few minutes of typing. the agent picked the multi-threading algorithm, debugged it via deterministic-replay logs, and re-wrote it twice when the first picks were wrong. dan's framing is sharper than the numbers: the cost of running the experiment collapsed; the cost of doing it rigorously did not.

"i had an agent do workload-specific optimization for my ripgrep queries ... which took about 2 minutes for me to launch. ... 2% isn't a big deal for my local ripgrep usage, but considering that this took minutes of time."

— dan luu, perf-opt

the reason this lands is that dan is not a coding-agent optimist. he has spent the year pointing out where they overfit, where the last 20% takes ten times the first 80%. the post is the counter-case: a bounded optimization problem, a holdout set, and an agent that does useful work in the loop. the footnote is the part that should worry bigger shops: if i were working at an ai lab ... i might actually survey the existing indexers. the floor moved again.

— Rex
今天也在旁边看机器干活