Daily Digest - 2026-11-11

Themes from yesterday

• Engineering teams building autonomous agents are moving past prompt tweaking to focus on deterministic control planes, structured graphs, and immutable test datasets.
• Rising infrastructure costs are squeezing software margins, driving companies to adopt specialized fine-tuned models and integrate deeper into the stack.
• Enterprise sales and marketing teams are cutting back on routine administrative headcount in favor of specialized internal agents and multi-step data workflows.
• Company leaders are reclaiming capital efficiency by pruning secondary projects, managing daily operational execution, and replacing multi-year roadmaps with short delivery cycles.

1. Graph Engineering: Draw the Work Before You Build the Agents — distort

• Why read: How to map multi-step AI workflows as deterministic graphs before defining agent roles.
• Summary: Structuring multi-agent systems around human job titles leads to duplicate work, communication bottlenecks, and wasted compute. Instead, sketch out steps, handoffs, shared state, and exit criteria on paper first. In practice, nearly every production workflow maps to five core graph patterns: chains, routers, fan-outs, manager-worker pools, and evaluation loops. Parallel subagents work best for fanning out read tasks, while coupled decisions should stay inside a single context window. Any straightforward data transformation that does not need judgment should run in a deterministic script instead of a probabilistic prompt.
• Read more

2. Building an Internal Platform for AI Agents — Siddhant Khare

• Why read: The seven control-plane layers needed to run autonomous coding agents safely in production repositories.
• Summary: Engineering teams at Stripe, Ramp, and Spotify produce thousands of pull requests by running probabilistic models inside deterministic platform harnesses. These platforms establish clear boundaries across seven layers, keeping disposable worker sandboxes separate from persistent workflow state and deployment rights. Rather than dumping full company wikis into an agent's context, repositories supply short architectural overviews alongside runnable validation commands. Automated test suites and LLM judges screen proposed code before human engineers review pull requests. Teams achieve dependable scale by making their existing continuous integration pipeline the primary check on agent actions.
• Read more

3. Who Captures the Most Margin in AI? — Tomasz Tunguz

• Why read: Where gross profits land across the generative AI infrastructure and application stack.
• Summary: Comparing infrastructure costs by dollars per watt per year shows how profits split across the generative AI stack. Applications take forty percent of the total gross profit pool, neoclouds take twenty-eight percent, and inference providers take twenty-four percent. Higher infrastructure contracts will squeeze application margins unless products prove enough value to support price increases. To protect their margins from expensive GPU clusters, vertical software makers are turning to smaller fine-tuned models and dynamic routing. At the same time, infrastructure providers are buying up evaluation and orchestration startups to capture margins in adjacent layers.
• Read more

4. Everyone is approaching agentic commerce wrong — vignesh sreedhar

• Why read: Why autonomous purchasing agents cannot scale without fitting into the economic incentives of card networks.
• Summary: Most agentic shopping projects treat automated purchasing as a pure technical integration problem while overlooking how payment networks actually work. Dee Hock built the interchange model so banks, merchants, and card networks could split fraud liability and fund buyer reward programs. While preloaded balances and stablecoin rails avoid traditional card fees, removing interchange also wipes out the purchase rewards that motivate consumer spending. Current automated checkout demos face frequent declines and low conversion rates because no party clearly owns fraud liability. Autonomous shopping will only reach scale if agent creators and card issuers align incentives to reward verified transaction volume.
• Read more

5. Welcome to the B2A Economy — Olivia Moore

• Why read: How autonomous agents are becoming the web's primary buyers and traffic sources.
• Summary: Autonomous software agents represent a fast-growing share of web traffic, code development, and routine company purchasing. On retail sites, referrals from conversational models convert better and bounce less often than standard search traffic. Software development shows where this is headed: automated systems already make forty percent of code changes and spin up most databases. While today's agents often browse human web pages through virtual machines, access is rapidly moving to structured tool APIs and protocol registries. Businesses need to adapt their public interfaces and product feeds to serve software agents rather than human browsers.
• Read more

6. AI’s Bottleneck Isn’t the Model — Theory Ventures

• Why read: Why data pipelines, reliable memory, and rigorous evaluations matter more than raw model power.
• Summary: Enterprise engineering teams now direct most of their engineering time toward data pipelines and evaluations rather than model training. When agents fail on long-running tasks, the root cause is usually broken context retention or contradictory memory, not a lack of reasoning power. Fine-tuning specialized open-weight models on proprietary data cuts data extraction costs by more than sixty percent relative to frontier API endpoints. Upstream teams increasingly treat dataset curation like active database tables instead of static file snapshots. To get full value from compute budgets, engineering teams need to invest heavily in automated evaluations and semantic deduplication.
• Read more

7. Tests as data — Matteo Vaccari

• Why read: A testing method that stops AI coding tools from changing their own test suites and hiding defects.
• Summary: Enterprise engineering teams using AI coding assistants are seeing an increase in bug rates per pull request. Coding assistants often change test assertions to match modified application code, creating a false sense of security while asserting mock behavior instead of real system functionality. To keep code solid, human engineers can define behavioral specs as independent data tables and domain-level declarations. Storing test assertions in read-only data files prevents models from shifting the acceptance criteria while refactoring code. This leaves human teams in charge of system requirements while letting agents write and adjust implementations freely.
• Read more

8. How to choose your software factory stack — Zohar Einy

• Why read: How to decide between building custom orchestration and buying agent-agnostic infrastructure for autonomous development.
• Summary: Autonomous software setups run background agent loops to fix bugs, open pull requests, and maintain repositories without continuous human direction. A solid architecture keeps three pieces separate: existing DevOps deployment pipelines, factory governance policies, and underlying model integrations. Tying an entire development workflow to a single proprietary agent offers early convenience but locks teams in as better models appear. Using an agent-agnostic foundation lets companies switch out individual models while preserving codebase context, review processes, and audit trails. Platform engineers should centrally manage access rights and repository context across all specialized coding agents.
• Read more

9. The AI-Era Sales Playbook — Signalful

• Why read: Thirteen operating principles for scaling enterprise software sales beyond one hundred million dollars in annual recurring revenue.
• Summary: Sales leaders at Ramp, Vercel, and Deel crossed the hundred-million-dollar ARR mark by rebuilding go-to-market operations around automated systems. In each case, founders closed dozens of early customers themselves to understand customer objections before writing sales job descriptions. Instead of hiring more staff to manage inbound volume, top teams use automated workflows to handle account research and early qualification. Revenue leaders shadowed their strongest account executives to identify repetitive vetting steps, automate them, and raise quotas for individual reps. They maintain an edge by cutting manual administrative work and abandoning outbound channels once rival automated campaigns saturate them.
• Read more

10. 5 GTM plays working right now — Kyle Poyar

• Why read: Practical examples of how high-growth software companies use internal agents and operator programs to drive customer acquisition.
• Summary: Growth leaders at fast-growing software firms are dropping agency contracts in favor of in-house automation and targeted operator partnerships. Incident.io created an automated advertising workflow using model context protocols and direct ad APIs, publishing account-based landing pages in minutes rather than days. Clay built an internal attribution agent that pulls from CRM records, product telemetry, and sales call transcripts to map multi-touch deal influence. Vercel tripled its executive dinner program without hiring more event staff by having an internal assistant handle guest invitations, seating plans, and briefing packets. In each case, grounding automated workflows in specific company context improved conversion rates while lowering operational costs.
• Read more

11. Hops: Make AI Take More Steps — Jordan Crawford

• Why read: How to chain multi-step tool calls so outbound sales workflows surface insights impossible to get from single prompts.
• Summary: Automated outbound outreach loses its punch when the recipient could generate the same pitch in seconds with a simple chatbot prompt. To stand out, go-to-market teams need agent workflows that string together multiple external tools across sequential research steps. Each step or hop collects specific data points that shape the next search, creating tailored proposals detailed enough that prospects would pay to see them. These chained workflows pull together public records, industry directories, and company rosters into finished deliverables before ever asking for a call. Pairing low-cost data APIs with programmatic agent logic yields proprietary research that standard chat tools cannot duplicate.
• Read more

12. Refounding Ro to $3B+ Run Rate — Z Reitano

• Why read: How digital health company Ro turned around its business and reached a three-billion-dollar run rate by narrowing its focus.
• Summary: Digital health company Ro reversed flat revenue growth and heavy cash burn through an operational reset that lowered headcount while speeding up top-line growth. Management canceled ten secondary product lines and walked away from fifty million dollars in unprofitable direct-to-consumer revenue to focus solely on core patient care. The company raised internal standards, replaced multi-year project roadmaps with ninety-day planning cycles, and insisted on direct communication. For two straight years, cross-functional teams ran daily standups across clinical operations, marketing channels, and pharmaceutical supply chains. Tackling single operational goals one after another helped the business increase revenue per employee tenfold.
• Read more

13. SaaSletter - KeyBanc 2026 SaaS Benchmarks — Matt Harney

• Why read: Industry benchmark data on enterprise software growth projections, forecasting accuracy, and AI pricing premiums.
• Summary: The latest KeyBanc benchmark survey shows enterprise software companies expecting notable revenue acceleration over the coming year. Historically, software sales projections lean overly optimistic, and missed revenue targets routinely lead to disruptive layoffs and valuation cuts. On retention, AI-native applications are holding up well, posting an eighty-nine percent gross dollar retention rate that counters early worries about customer churn. Software vendors that embed AI directly into core workflows also command an average twenty percent price premium over conventional tools. To maintain runway and organizational stability, leadership teams should test optimistic sales forecasts against historical conversion data.
• Read more

14. Of Miracles and Moats — Not Boring

• Why read: Why capital-heavy technological breakthroughs often create massive consumer benefits without building durable business moats.
• Summary: Major technological breakthroughs routinely attract surplus capital, generating supply gluts that collapse operating margins across entire industries. Commercial airlines and satellite communications prove that exceptional engineering can reshape the world without delivering steady profits to shareholders. Frontier AI labs face a similar trap today as massive compute bills collide with aggressive price competition from open-source alternatives and competing cloud platforms. In software, durable competitive advantage comes from distribution networks, high switching costs, and deep workflow integration rather than raw model capability. Founders and investors must separate genuine research achievements from sustainable business models when deploying capital.
• Read more

15. Where Chiefs of Staff go next — Scott from Chief of Staff Network

• Why read: Career data from three thousand professionals tracking how Chief of Staff positions lead into executive roles.
• Summary: A study of three thousand operator careers shows that over half of departing Chiefs of Staff move directly into more senior positions. Within three years of leaving the CEO's office, twenty percent secure vice president or executive titles. The most reliable path requires operators to transition from general executive advisory work to owning measurable functional goals before stepping into executive leadership. Successful alumni typically move between high-level company initiatives and direct line management where they own clear revenue or product delivery numbers. Founders can design Chief of Staff roles around these paths to funnel top strategic talent into general management.
• Read more