|
Hey there 👋
Happy Fourth. If you’re reading this between the grill and the fireworks, I’ll keep it to one idea.
Last year I spent an afternoon hunting for a cheaper model to shrink an agent’s bill. Wrong hunt. The bill was not the model — it was that we shoved all ninety tool definitions into context on every single call. We were paying to remind the agent what it could do, over and over.
This week two research teams put hard numbers on that exact mistake. The cheapest win in 2026 is not a cheaper model. It is not sending the tokens you never needed to send.
Let’s walk it.
The Big Thing
Alibaba cut an agent’s token bill 99% by not loading every tool
Here’s the number that should make you re-open your cost dashboard. Alibaba researchers built SkillWeaver, a framework that took one complex, multi-step task from about 884,000 context tokens down to roughly 1,160 — a 99%-plus cut — while keeping tool-routing accuracy near 92%.
How it works is the whole lesson. Most agents with a big tool library do the naive thing: dump every tool definition into context, then ask the model to pick. SkillWeaver instead decomposes the task first, maps the sub-steps and their dependencies as a graph, and retrieves only the handful of tools each step actually needs. The model never sees the other 800 tools. You stop paying to remind it what it can do.
The third-thought read: the headline invites you to file this under “Alibaba beat everyone,” and that misses the point. The lesson is about your own bill. For a lot of production agents, the biggest line item is not the model’s per-token price — it is orchestration waste: giant system prompts, every tool loaded every call, memory blobs re-sent turn after turn. Chase a cheaper model and you save 20%. Fix how you load tools and you save an order of magnitude.
And you don’t need Alibaba’s code to do it. The pattern is portable: plan the task, retrieve tools per step, keep the catalog out of context. If your agent talks to more than a couple dozen tools, this is the holiday-weekend audit — pull one expensive workflow and count how many tool tokens ride along on every call that never uses them.
The honest caveat: 884K-to-1,160 is one reported task on Alibaba’s own benchmark, and 92% routing accuracy means the router still misses roughly one call in twelve — so guard the steps that matter. Treat the exact number as a ceiling, not a promise. The direction is the deployable part.
Ship it? The framework — watch it; it is research, not a product you install Monday. The pattern — adopt it now. Auditing tool-loading is free, reversible, and the fastest money you’ll save on an agent this quarter. VentureBeat write-up · method paper.
Sources: VentureBeat, Dataconomy, arXiv — figures are Alibaba’s reported benchmark.
Tour de Headlines
🧭 Zoom is buying its way into the buyer-signal layer. Zoom signed a definitive deal on July 2 to acquire Common Room, an AI-native go-to-market platform whose RoomieAI agents run account research, message personalization, and prospecting. It folds into Zoom Revenue Accelerator; terms were not disclosed and the deal closes in the coming weeks. The tell is Common Room’s customer list — Anthropic, Okta, Snowflake, Notion, Autodesk. The third-thought: incumbents are not buying agents, they are buying the proprietary buyer signal the agents run on. For RevOps, the question stops being “which agent” and becomes “who owns my intent data after this M&A wave.” Zoom’s announcement.
🧰 Someone built Vercel for MCP servers. Manufact — YC S25, formerly the open-source mcp-use SDK (5M+ downloads) — launched MCP Cloud on July 2: push an MCP server from a GitHub commit to production in under 60 seconds, with deploy, scaling, auth, access control, and observability handled for you. One codebase serves ChatGPT Apps, Claude Connectors, Gemini, Cursor, and Claude Code. The builder read: the MCP gold rush is shifting from “write a server” to “operate a fleet of them.” If you’ve duct-taped MCP servers into prod, the 3-hour project is deciding whether an ops layer beats your homegrown glue — before the next MCP spec on July 28 multiplies what you maintain. Via VentureBeat.
🛡️ A big integrator put a frontier model to work on defense. Cognizant and OpenAI launched a service on July 2 that applies GPT-5.5 with Trusted Access for Cyber to defensive security work — secure code review, threat modeling, vulnerability triage and validation, detection engineering, incident response — with human validation at every step and scoped, monitored access. Cognizant is running it internally first, as “Client Zero,” backed by 5,000+ security staff. The point for builders is not the model; it is the shape: a governed template that embeds into workflows you already run and augments deterministic controls rather than replacing them. If you’re weighing an agent in your SOC, this is a reference design to copy. The release.
|
Sponsor
You audit what your agents waste. What about your calls?
This whole issue is about cutting the tokens your agents never needed to send. Your sales and CS calls carry their own waste — the misread on discovery, the deal thread where two people “agreed” to different things, the rapport that quietly never formed. It completes without a warning. RapportScore scores how your team actually connects — on calls, in email, across every deal — so the misfires surface before they cost you the quarter. You instrument your agents. Instrument your people too.
See your team’s score →
|
Tool of the Day
🎯 Profound Aim
an always-on background agent that turns AI-search signals into marketing work — and hands the work to other agents.
Profound launched Aim on July 2, calling it the first always-on background agent for marketing. Here is the loop it runs: it watches your brand’s citations, sentiment, and mentions across AI answers; when something moves — say, mentions drop — it diagnoses the likely cause, writes a memo, spins up a project with tasks, and routes execution to specialized sub-agents, keeping a human on every approval. Then it measures whether the finished work actually moved the metric, and feeds that back into the next recommendation.
Even if you never touch Profound, steal the shape. This is the cleanest example this week of the manager-agent pattern: detect → brief → delegate → approve → measure. That loop is what B2B teams will copy far beyond marketing — into support, RevOps, security triage. (Worth noting: Profound’s $96M round was back in February; the news here is the product, not the funding.)
See how Aim works
Worth a Click
- The other “don’t send tokens” win. Microsoft Research introduced Memora (ICML 2026), a memory layer that splits what an agent stores from how it retrieves — rich content, lightweight cues — and reports up to 98% fewer context tokens while setting state-of-the-art on long-memory benchmarks (LoCoMo, LongMemEval). Put it beside SkillWeaver and the theme is loud: 2026’s biggest cost wins come from sending fewer tokens, not buying cheaper models. Research, not a product — but the code is on GitHub. Microsoft Research.
- Frontier models are shipping behind a gate. OpenAI’s next family — GPT-5.6 Sol / Terra / Luna, with a “Sol” tier built for the hardest agentic work — is previewing to only ~20 government-vetted partners, with general availability promised “in the coming weeks.” Same window, OpenAI floated donating 5% of its equity to a U.S. sovereign-wealth fund (early, preliminary talks). The builder takeaway is not the benchmark; it is that access to the best agent models now arrives in windows, not launches — plan procurement around them. Via VentureBeat.
- Where the money went this week. The checks favored evaluation and infrastructure over another chat UI: Talp raised a $20M pre-seed (Formus Capital, a16z Scout Fund) for AI personas that simulate customer behavior to test products and pricing before launch; Manufact took $6.3M for its MCP deploy layer (above). Read the pattern: investors are backing the layers that make agents testable and deployable, not the demos. TechFundingNews.
Two research teams, same lesson, one week apart. SkillWeaver stopped loading every tool. Memora stopped re-sending every memory. Both cut cost by 90-plus percent — and neither one swapped in a cheaper model. So before you spend the quarter benchmarking models to shave your agent bill, spend three hours this weekend on the boring audit: what are you sending on every call that the task never needed? The catalog of 800 tools. The 4,000-token system prompt. The whole conversation history, re-sent turn after turn. The cheapest agent win in 2026 is not a cheaper model. It is the tokens you stop sending. Happy Fourth — go find yours.
See you tomorrow. — The Agent Stack
Built for people who ship AI, not people who tweet about it. Published daily mornings by Pixiu Media Holdings LLC.
|
