# Decision Points for Embedding AI Agents in Enterprise Systems # Employee-facing vs Customer-facing 🎯 The Hook Are you using the same design for your internal employee agent and your customer-facing agent? The trust model, data boundaries, guardrail strength, and blast radius of failures are fundamentally different depending on who is using the agent. Get this fork wrong early and you will either expose customer data through a lax design or crush employee productivity with excessive restrictions 🔑 📋 Overview Whether the agent's users are employees bound by employment contracts and NDAs or external customers who may submit adversarial inputs is the very first architectural fork. For employee-facing agents, you can assume a baseline of trust — authentication is unified through corporate IdP (Okta / Entra ID), data stays within internal knowledge bases (Notion / Confluence / Box) and business systems (Salesforce / ServiceNow / Workday), and the failure cost is typically "rework" rather than "lawsuit." Customer-facing agents are a different world: you must design for jailbreak attempts and indirect prompt injection, outputs directly affect brand reputation and legal liability, strict per-tenant data isolation is mandatory, and the authentication layer must handle CIAM (Auth0, etc.) including anonymous access. Scale is also orders of magnitude apart — thousands of employees versus millions of customers with 24/365 spike resilience 📊 🔍 Decision Points The two driving variables are user trust level and failure blast radius. Users are employees only → Employee-facing design is sufficient External customers are included → Customer-facing design is mandatory Both are included → Design as separate planes with isolated data paths The critical insight is distinguishing what can be shared from what must be isolated. The orchestration layer and model gateway (AI Gateway) can be shared across both planes. Data access paths, guardrails, and audit logs must be separated. For customer-facing output pipelines, integrate DLP (Data Loss Prevention) to prevent internal data leakage ⚡ 💡 Key Details Employee-facing characteristics: - Input trust assumption holds (employment contracts and NDAs provide accountability) - Data scope is limited to internal knowledge bases and business systems - Failure cost is "inefficiency" or "rework" — mostly reversible - Concurrent users in the thousands to tens of thousands, concentrated during business hours Customer-facing characteristics are fundamentally different: - Design must assume adversarial inputs (jailbreak, indirect injection) - Incorrect answers or data leaks escalate to legal liability and brand damage - Strict per-tenant data isolation is non-negotiable - Concurrent users in the tens of thousands to millions, 24/365 spike resilience required In hybrid configurations, a typical pattern is an e-commerce platform on Shopify where the customer-facing inquiry agent and the internal order-operations agent share the same orchestration backbone but run on separate planes. On Zendesk, the customer-facing agent is given access only to a read model projection of publicly shareable knowledge, with direct access to internal Notion and Box data completely blocked 🔒 ⚖️ Trade-offs Reusing an employee-facing design for customers is the most common and most dangerous mistake. Data access scopes that are acceptable internally get applied to customer-facing agents, and suddenly one tenant's data leaks to another. Going the other direction — applying strict customer-facing guardrails company-wide — tanks employee productivity and kills the ROI of the entire agent initiative 😩 Deferring plane separation to "later" is equally risky. Starting with a single stack and planning to separate at customer launch is a recipe for ballooning migration costs. Separation must be decided at initial design time. Audit log co-mingling is another overlooked trap — when customer PII and internal data share the same log store, data retention policies and access controls become unmanageable ⚠️ 🛠️ Use Cases E-commerce operations + customer support: A Shopify-integrated retailer runs a customer-facing inquiry agent (strict guardrails, tenant isolation, DLP enforcement) alongside an internal order-management agent (broad internal data access, wider operation permissions) on the same orchestration backbone. The model gateway is shared; only the frontend and data paths diverge 🛒 IT helpdesk: An employee-facing IT support agent has broad access to Active Directory, Jira, and Confluence, automating password resets and software provisioning. When the same platform is extended to customer-facing support, access is restricted to the public knowledge base, guardrails are significantly hardened, and topic restrictions, tone control, and refusal policies are layered on 📚 Practical tip: If there is even a possibility that you will need both employee-facing and customer-facing agents, design for plane separation from day one. Retrofitting separation is among the most expensive categories of technical debt 💪 #AIAgents# #EnterpriseArchitecture#

Autopilot Update: I have reviewed the positions and there's nothing I feel like should be changed. We were well diversified enough that we handled recent drag with minimal downside, and are now positioned for what may be a decent leg up. Repricing is done from yesterday. The hike was priced in, not Warsh's tone or pressure, so that down and bounce repricing is c o m p l e t e. Oil price is falling. This means, despite the dot plot on the SEP, the next likely rate hike date (Dec) might lose traction if pass-through inflation dies out. Granted, we have yet to see much of the sustained oil prices it hit economic data in my opinion, but it would be lagging. So the market will be guessing. AI trades are running. Quad witching tomorrow. So: hopeful that this sustained period of stagnation and consolidation in prices is being broken out of, but we will have to see if it's a relief rally or a run with sustainable legs. Time will tell, as always. Adam

I think Muse is what AI is supposed to look like. Feels like every lab has been shipping the same thing recently with minor incramental improvements. Text boxes in front of a model = shiny new toy syndrome! No one cares if Muse Spark is the best model in the world rn because it literally doesn't matter to 99.9% of people. Only the top 1% of coders/engineers care about what's changed from one Claude release to the next ChatGPT release. Personally, I have found little improvement since Fable 5 was released and I think that says a lot. I'd probably say it's regressed recently for higher end research tasks. Muse makes me more convinced now that Anthropic/OAI have absolutely zero moat at the consumer level. Enterprises are a different story and that's where both frontier labs need to (already are) focus rather than doomerism and frontier pacing bs. Agents are where AI endgame probably goes - most users just care about usefulness so having one run laborious tasks is way more important than a tiny push on the frontier by Fable 5.1 / Astra to most people. Checking in with your agent will become force-of-habit, same way you check your WhatsApp or emails every morning. Not everyone logs onto Claude or ChatGPT in the same way. (Already a known thesis on Agents, but yeah, worth resurfacing). For the average Joe, this is all the matters. Same way most people don't change phones from one generation to the next because their current phone is good enough. Same way laptops were so important for the PC industry because they were just more convenient than desktops. In that sense, I guess you could say that Muse has pushed the frontier from a usefulness POV. I think we think that most people care about AI like we do...they don't lol. If AI is gonna become mass adopted, it'll be via agents doing stuff for people just how Muse is rn. Since most people already know ChatGPT (noun for AI e.g. "ChatGPT it" similar to "Google it"), it seems like OAI doing a solid enough agent could be enough to oust Anthropic longer term at the consumer layer to have all your context with one lab in an agent + frontier model combo. Sorry for the ramblings and no idea if this logic holds lol - I'm walking to get lunch and wanted to write after a stressful morning!

BRASILIA E AÍ 🧡♾️ Tonight’s the last show in South America for The Lifetimes Tour and I’m so full of joy and many varieties of empanadas. You know when I say I love you for life, I mean it. Thank you for showing up like you did Cats (& Rats too… I guess) ♥️

AI is still so early, but so bullish. We're witnessing a huge unlock of economic value with the adoption of AI models like Claude, and more recently, agents like Grok Bot. This then filters down into the most bullish of bull points for AI. Enterprise AI adoption increases productivity -> increases enterprise earnings -> increases token demand as a flywheel effect -> increases ROI for the AI labs -> increases demand for AI infra e.g. compute. I recently heard from an ex-colleague that a large consulting firm is increasing their AI spend by 100x at their London office for their back office teams. Just incredible growth really. We can kinda see the enterprise flywheel manifest in the $NVDA earnings too. AI Clouds, Industrial & Enterprise revenue was up 25% sequentially compared to 13% for the hyperscalers. I think that this is also why $NBIS and $CRWV keep signing enterprise and lab capacity faster than capacity comes online too. 2027 capex will be huge because the labs / hyperscalers still can't get enough compute to serve enterprise demand right now...and that's before AI adoption has scaled properly across the wider economy. Fun times ahead with AI.

Naomi Watts is the voice of Peter's new AI assistant E.V. in ‘SPIDER-MAN: BRAND NEW DAY’. They previously starred together in Tom Holland's first major movie ‘THE IMPOSSIBLE’.

Anyone who ever owned a foil card knows the tilt. Now you can just describe one. We asked for a fruit fly playing with a Rubik's cube on a starry background. Back came "Fruit Fly AI" — ukiyo-e composition, coloured sumi-e linework, a two-state morphing front that flips. It opens in a browser. Drag to rotate, flip to the back, tilt it and the foil moves like cardboard does. $1 a card. Under an hour.

as we keep moving towards personal software & owning the tokenflow, the first thing to be tackled is the "personal harness" - implemented as "ai inside an app" vs "app inside ai" (i.e. the inversion of "codex has browser pane" - "app in your browser has codex inside of it") fx by vercel is act 1

Enterprise vibecoding is a shit show. Most companies subject themselves to one of two failure modes: • Blindly enable vibe coding. It becomes a matter of when (not if) some data security bomb will drop. • Blindly deny vibe coding. The company deprives itself of decentralizing problem solving, and shadow AI (I.e people using personal accounts runs rampant). Which is why customers/prospects ask for some silver bullet to responsible, enterprise-grade vibecoding weekly. The typical question we get: “How can we let non-technical employees build production software without a data security nightmare or millions in wasted tokens?” And now having solved this problem for a number of companies including a multi billion dollar PE firm and F500 retail biz, we have a proven framework you can use within your business. First, why is vibecoding good for business? • Employees are less likely to solve the problems of the business and are more likely to solve their own problems (which in turn as adoption grows, solves parts of a business). • Employees / vibe coders are going to build anyway. Sanctioned / safe / governed tools > shadow IT/AI. • SMEs are the ones who have the deep knowledge on an industry/workflow, not the engineer. Empower the SMEs to build, and they can build useful tools. • 95/5: most apps will fail, your winners gives your business outsized returns. Okay, now let’s assume you’re bought into vibecoding…. Second, you have to be clear about what you’re optimizing for when enabling enterprise vibecoding. 1. Turning SMEs into problem solvers A critique of engineers as old as time is that they don’t sit close enough to the work to truly understand the problem/end user. Whether you agree or not, it’s not a hot take to say that subject matter experts doing the work deeply understand the problems and friction points that need to be solved through product. 2. Enabling multiplayer AI Building an app that runs locally on your computer and solves a clear problem is good. Getting that app to be used & improved by everyone in your department that experiences the same problem is how you get far more leverage. 3. Allowing IT to sleep at night If customer data leaks or an api key is exposed, guess whose ass is on the line. IT. No wonder they’re so paranoid about things like company-wide vibecoding. They’re not anti-vibecoding, though. They’re anti-having the right guardrails in place to prevent the worst day of their careers. 4. Making data access programmatic Pre-AI, IT teams could get 5 data requests a week. Today, that number could be 50/day. As data access goes parabolic because it’s the foundation for performant AI, tools for IT must evolve in-kind. 5. Production-grade by design For product/engineering to take vibecoding seriously, enterprise consistency has to be built into what employees build. If at any point someone has to take over, or review, they are shaped relatively the same, adhere to the eng org’s conventions, and are easy to understand. Now, how do you put this into practice? Third, you have to build a software pipeline that enables secure & responsible, production-grade vibe-coding at scale. We call it the Citizen SDLC. Step 1 is Discovery. Input: An employee works with AI to described their idea, the problem, users, data needs, and desired behavior. Output: This gets turned into structured product requirements doc. Step 2 is Request. Input: PRD becomes one structured request to IT with a named owner, users, data classification, integrations, and scope. Output: Every app enters IT triage on the record before any build work begins. Step 3 is Triage. Input: AI classifies the request along two axes. • Shape: what kind of app is it? artifact generators, workflow automations, CRUD apps, or interactive dashboards. • Blast radius: how much damage could this build do if it went sideways? Output: Three routes out. • Approved: blast radius inside every threshold, complete brief, high confidence. • Reuse: overlaps an app that already exists, so the requester gets routed to that app’s owner • Escalated: A dimension crosses its threshold. IT gets the full request. Step 4 is Provisioning. Input: One approval, the paved road for the app’s shape (architecture, identities, governance, etc), app metadata and owner. Output: platform stamps the app from the road for its shape: a repo, firm sign-in, a deploy identity, a private environment, and its own database, all defined as infrastructure-as-code that IT owns and versions. Step 5 is The Build. Input: The employ prompts their coding agent (Claude Code, Codex, etc) inside the scaffolded architecture, adds product behavior, and opens a PR when ready to publish. Output: Finished app. Everything an engineer would normally carry is carried by the rails instead. Step 6 is Run & Change. Input: New features or app changes made by the original owner or collaborators in the coding agent. Output: Humans do not review routine changes. The checks are the review. What reaches a human is the consequential tail, detected mechanically: destructive schema changes, new dependencies, etc. Hopefully you found this approach to vibecoding at your company helpful! If you want the full breakdown of the six step Citizen SDLC process in excruciating detail, check this link. It’s an end-to-end playbook for responsible, production-grade enterprise vibecoding: