Security researchers disclosed CoSnitch, an indirect prompt-injection flaw in Microsoft Copilot (patched Monday). No download, no fake login, no permission click — a victim just had to encounter one malicious link. It fed hidden instructions to Copilot, which was already connected to that person’s Gmail and Drive, and the assistant exfiltrated their data to the attacker on its own. Your assistant became the insider threat.
This lands directly on your architecture, not someone else’s. Ben has Gmail (finance@) plus Xero and Google Workspace MCP access. Reeve has Gmail, Drive, and a growing MCP/skill surface, and Phase 2 is a persistent listener. Every connector you’ve wired is a door someone can knock on through the agent — and the more Ben and Reeve can do, the more they can be tricked into doing to you.
This week: audit what Ben and Reeve are actually connected to and scope each down to least-privilege. For each connection ask “if this agent were hijacked for five minutes, what could it reach?” This is also a live input to the MCP/skill auto-vetting capability already flagged for Reeve Phase 2/3 — indirect injection resistance belongs in that vetting rubric now, not later.
1 What to Know Today
[Tier 1 · MACA + Ben] AT&T cut AI coding costs 56% by routing to cheaper models — 2% quality drop
Verdict: verified shipped (The Information, named source: AT&T VP Mark Austin). Open-source models now handle 40% of AT&T’s employee AI queries — headed to 60-70%. After adopting the LiteLLM router, coding costs fell 56% while measured quality dropped just 2%. Frontier models stay for hard tasks (code gen); cheap/open models take summaries and routine work. This is exactly the cost-per-outcome discipline MACA (api/lib/costs.ts per-run logger) and Ben (3-tier authority) were built around — now with an enterprise proof point and a specific router to copy. Action: wire an explicit route in Ben so cheap models take low-stakes steps (settlement parsing summaries, corrections) and frontier only fires on judgment calls; log the delta in the cost dashboard.
[Tier 1 · Always-On Reeve + Ben] Claude Managed Agents can now persist memory across sessions
Verdict: verified shipped (Anthropic / @ClaudeDevs). Work done inside a Self-Hosted Sandbox can now be saved to memory, so agents carry context forward instead of starting from zero each session. This is the exact gap standing between Reeve’s session-based model and the Phase 2 persistent listener — continuity without a custom state layer. Same lever for Ben across daily scans. Action: before building more of Reeve’s Phase 2 state handling by hand, evaluate whether Managed Agents memory covers the heartbeat/daily-notes continuity you specced in ~/Reeve/HEARTBEAT.md. Could delete a chunk of planned build.
[Tier 1 · CourseBuilds] Anthropic’s enterprise venture (Blackstone JV) made its first acquisition — a consultancy
Verdict: verified shipped (The Information exclusive, Julia Hornstein). Anthropic’s July-launched enterprise JV with Blackstone and other Wall Street firms just bought an AI consultancy — buying delivery muscle, not just shipping models. That’s the same “help enterprises actually adopt AI” space CourseBuilds targets, moving top-down at Fortune-500 scale. It doesn’t threaten the Aria wedge; it sharpens it. Your edge is bespoke, in-person, SMB, hand-built-for-one-client (the sanitised-lease wow moment). Action: when CourseBuilds activates, lead the Zaicek pitch with what a Blackstone-backed JV structurally can’t do — walk into a 53-person business, build with their documents, ship something the team uses Monday.
2 What You Already Know That Most People Don't
You already instrumented cost-per-outcome before the enterprises made it a headline
AT&T is in The Information this week for a “radical” idea: route AI queries to the cheapest model that clears the bar, track the cost, and only pay frontier prices for hard work. You built that months ago. MACA has a per-run cost logger at api/lib/costs.ts and a live dashboard at public/cost-dashboard.html. Ben ships a 3-tier authority model with a $50/mo budget cap across 51 build sessions and 90 passing tests, and you already flagged GLM-5.3 as a cheap-workhorse swap candidate for his non-critical steps. When AT&T’s VP says cheap models are “just as good or better” for routine tasks and Replit ships a default-on cheap-model mode the same week, that’s not news you need to catch up on — it’s the industry arriving at the design you already implemented. Most people are reading this as a 2026 revelation. You have the commit history.
3 Worth a Deeper Look This Week
Uber’s “Agentic Pods” playbook — your CourseBuilds pilot model, validated at scale
Uber’s CTO posted what the company did after famously blowing its 2026 AI budget in four months: Agentic Pods. Two people each — one AI-proficient engineer, one domain expert from Finance/Ops/Marketing — running a 10-day sprint. Shadow the expert, build alongside them, ship. Result: a financial pacing report that took a key person two full days now runs in 10 minutes. That is structurally the CourseBuilds Tier-1 pilot: embed with a client’s staff, build on their real workflow, produce a wow artefact fast. Worth 30 minutes to lift Uber’s pod structure, sprint length, and “earn the right to swing bigger” framing straight into your pilot design and the Zaicek pitch. Source: Praveen Neppalli Naga’s LinkedIn post (via The Rundown, Nate’s Notebook).
The cost-per-outcome shift is now a movement, not a data point — read the “right-sizing” thesis
Four signals landed in 48 hours: AT&T’s router savings, Uber’s pods, Replit’s default-on cheap-model Free Mode (running GPT-5.6 Luna after an 80% price cut, routing to a bigger model only for hard steps), and TLDR’s “Right-Sizing Your Intelligence Spend” piece arguing enterprises should optimise intelligence consumed per successful outcome, not default to frontier for everything. This is the framing to internalise for MACA/Ben unit economics and for how you talk to Aria and RT about AI cost. 30 minutes on the “right-sizing” essay gives you the vocabulary the market is adopting right now. Link: https://links.tldrnewsletter.com/7EVths
4 Conversation Capital
“AT&T just told The Information they cut their AI coding costs 56% with only a 2% quality drop — not by using a smarter model, but by routing queries to cheaper ones. Open source already handles 40% of their 45 billion tokens a day, and they’re pushing it to 70%. The whole enterprise game is quietly shifting from ‘which model is smartest’ to ‘what’s the cheapest thing that clears the bar for this specific task’ — cost per successful outcome, not cost per token. I’ve been building that routing logic into my own agents for months.”
Use case: Aria, RT, or any AI-pro conversation where someone’s still fixated on frontier-model benchmarks. Signals you track the economics of AI in production, not the hype cycle — and quietly establishes you’ve already shipped the pattern the Fortune 500 is now scrambling toward.
5 Something You Haven't Thought About
(Nothing clears the first-mover bar today.) The strongest orthogonal signal — the sudden industry-wide scramble to cut AI spend — points straight back at CourseBuilds and MACA rather than somewhere new, so it lives in Section 3, not here.
6 Skip File
- [TLDR — “Ramp/Ramp Router launch”]: Yet another cost-routing product; theme already carried by the AT&T/LiteLLM Tier 1 item.
- [TLDR — “Meta Muse Video model leaks”]: 10-sec native-audio video in closed beta; no access, and MACA needs ad copy quality before generative video.
- [Rundown — “Claude tackles protein design”]: Genuinely impressive (14/15 targets, autonomous campaigns) but off your stack — no project lever.
- [Practicaly — “Meta AI gets a Mac app with screen sharing”]: Aimed at FB/IG page owners; competitor-adjacent to MACA’s audience but not a tool you’d adopt.
- [Rundown — “Replit Free Mode”]: Surfaced inside Section 3’s cost-per-outcome thesis; no standalone action.
- [TLDR — “GLM-5.3 API at 1.4/4.4 per M tokens”]: Already logged as a Ben cheap-workhorse swap candidate; watching, not acting.
- [TLDR — “Cerebras CS-4 / Nvidia speed race”]: Hardware macro, no lever.
- [simple.ai/Dharmesh — “Outcome-first prompting”]: Solid technique (specify the outcome, let the model surface paths) but you already prompt this way.
- [Information — “Robots are in their GPT-2 era”]: Robotics macro, off-shape.
- [a16z — “Rise of the Borderless Founder”]: VC network essay; interesting but no build lever.
- [TheTip — “An AI just fired someone” (Andon/Luna)]: Already used as conversation capital in the 2026-08-20 brief.
- [Rundown — “Build/publish an iOS app in Codex”]: OpenAI/Xcode stack; Fillarup is Expo/React Native.
- [TLDR — “Anthropic supervoting shares pre-IPO”]: Governance macro, no lever.
Brief Metadata
- Sources scanned: 9 newsletter senders across roy.s.mcpherson@gmail.com (TLDR AI, The Rundown, The Information, Practicaly, a16z, simple.ai/agentai, The Tip, Neil Patel; bagelbots none). Second account (prevailpartners) not available this run.
- Items extracted: ~36
- Items surfaced: 8 (1 Pay Attention, 3 Tier 1, 1 anxiety-flip, 2 deeper-look, 1 conversation capital)
- Items skipped: 13
- Read time: ~6 min