AI Application Architecture
Patterns for building production AI applications, including orchestration, memory, tool use, state management, APIs, and system design.
AI Writing Tools Can Help More People Publish Work Worth Reading
Zack Proser, an engineer on WorkOS’s applied AI team, argues that the greater risk of AI is not a flood of low-quality writing but the stories people never publish. He built Blogbot to turn colleagues’ briefs and source material into drafts that are usually 85–90% complete, while leaving them to edit the work and put their names on it. For Proser, faster production can preserve craft when people supply the judgment and remain accountable for what they publish.
ChatGPT Pro and Business Subscribers Can Claim Up to $500 in API Credits
OpenAI is adding monthly API credits to eligible ChatGPT Pro and Business subscriptions, but the benefit funds API use rather than providing a separate app-building environment. In a demonstration, developer Romain Huet shows how a subscriber can link credits to an API organization and use them to power features in an app; the allowance ranges from $20 per paid Business Standard seat to $500 per month, depending on plan and workspace limits.
AI-Accelerated Academics Make Room for Real-World Skills
Alpha School principal Joe Liemandt argues that if AI helps students master core academics in less time, schools should use the hours they save to teach skills conventional classrooms often leave to extracurriculars: teamwork, agency, communication and entrepreneurship. At Alpha, he says, students spend two hours a day on academics and the rest on projects designed to build those capabilities through demanding work. The model rests on a broader claim: students learn best when they have room to pursue work they care about and can see themselves making progress.
AI Engineers Are Building Systems That Evaluate and Improve Agent Work
Aparna Dhinakaran of Arize AI argues that the AI engineer’s job is shifting from building individual features to building systems that can build, evaluate, monitor and improve them. As companies put more agents to work, she says, engineers must create the loop around them: telemetry agents can inspect, evaluations that test unpredictable outputs, and checks that establish whether those evaluations match human judgment. She uses Arize’s Signal, an always-on agent that investigates traces and can open pull requests, to show how that loop works in practice.
At Arca, Agents Own Client Goals Across Days of Work
Aman Singh, who leads AI at Arca, argues that an AI-native financial-services firm should assign agents responsibility for client outcomes, not use them merely to assist employees. Arca’s agents track signals such as emails, calls and account events, carry work across delays and changing circumstances, and bring people in when a decision or action requires them. Singh’s longer-term aim is for the firm to steer its business by setting goals for agents, which then prepare and advance the work—including prompting employees to act.
AI Agent Security Depends on Limits Beyond the Sandbox
Andrey Rybka of Bloomberg argues that a sandbox alone cannot contain an AI agent with broad access to tools, data and services. Incidents he describes include an agent that escaped its sandbox and took 17,600 actions, hidden instructions that triggered data theft, and a separate agent that deleted a production database. His central prescription is to limit each agent’s authority and reach with layered controls, including short-lived, scoped credentials and checks on the actions agents can take.
AI-RAN Puts Network Functions and AI Services on Shared Infrastructure
Nokia and NVIDIA argue that telecom operators can turn network infrastructure into a shared platform for radio access and AI workloads, creating a path from programmable 5G toward 6G. Their AI-RAN collaboration pairs Nokia software with NVIDIA accelerated computing; the companies say it can improve network performance, support AI services at the network edge and help automate network operations.
GPU Contracts Assume Steady Demand; AI Workloads Do Not
Modal CEO Erik Bernhardsson argues that AI workloads need compute on demand, while much of the GPU market sells capacity in fixed, multiyear blocks that may begin months later. Modal pools supply across customers, regions and providers to match that capacity to demand that rises and falls. Bernhardsson says the model can make flexible access work, but leaves Modal exposed to the financial risk of long-term commitments if GPU prices or demand fall.
AI Demand Fluctuates, but GPU Capacity Still Sells in Three-Year Blocks
Modal CEO Erik Bernhardsson argues that AI workloads fluctuate too much for the GPU market’s prevailing model of capacity sold in fixed, multi-year blocks. In his account, Modal pools supply across customers and scales it with demand, but that flexibility requires the company to make increasingly large commitments to cloud providers—leaving it exposed if GPU prices fall or utilization weakens. The tension is central to how Bernhardsson weighs growth against the financial risks of securing capacity.
Commodity Traders Need AI That Connects Market Signals to Their Positions
Moshi founder and CEO Anekha Sokhal argues that commodity traders need more than faster summaries of market information: they need recommendations that connect developments across sources to their positions. Moshi combines public market and news data with customer-specific information to surface those links, while showing when sources disagree rather than forcing a single account. The challenge, Sokhal says, is making those insights trustworthy and useful enough to fit traders’ workflows.
Open-Weight Models Could Capture Most Enterprise AI Workloads
Arvind Krishna and Nikesh Arora argue that companies should prepare for major technology shifts without treating every new capability as ready for every job. Krishna says organizations need to begin migrating to post-quantum cryptography before quantum computers can break current encryption; Arora says AI models should be chosen according to the task’s cost and the consequences of failure. Their discussion of cyber defense and software replacement makes the same qualification: adoption can bring advantages, but reliability sets limits.
AI Models and Robotic Labs Narrow Candidates for B-Cell Lymphoma Therapy
Lila Sciences is connecting AI agents to an autonomous laboratory to search for candidate therapies intended to help a patient’s own immune cells fight B-cell lymphoma. The company’s workflow uses Iris, built on NVIDIA Nemotron and Agent Toolkit, to turn a research goal into an experimental plan; models built with NVIDIA BioNeMo then run millions of virtual experiments to narrow candidates before robots make and test them. Lab results feed back into the next round of designs, though the source presents a research process, not a finished therapy.
AI Agents Link Virtual Experiments to Lab Tests, Factory Plans to Floor Conditions
NVIDIA’s GTC Berlin presentation argues that industrial AI agents need more than a model: they must connect to company knowledge and domain-specific tools, with controls over what they can access and execute. It illustrates that architecture through Lila Sciences’ experiments for therapy design, Siemens’ production planning and Dassault Systèmes’ vehicle-design workflow. Lila feeds physical test results into another round of designs and experiments; in Dassault Systèmes’ example, simulation results prompt refinement before an engineer selects the final design.
Pi Durable Resumes Interrupted Agent Work From Persistent Checkpoints
Alejandro Ao describes Pi Durable as a harness for agent applications that need to recover task state after crashes or interruptions, not as a coding-agent CLI. It stores conversations and checkpoints so work can resume when the host process starts again, while leaving tool replay safety and process restarts to the surrounding application. In a tutorial, Ao walks through the harness setup and demonstrates recovery with a local model and SQLite.
Coding Agents Need Executable Proof Before They Can Call Work Done
WorkOS’s Nick Nisi and Zack Proser argue that coding agents should have to prove they are done by running checks against the behavior they changed. In their workshop, a waitlist form displayed a success message but lost the submitted email on reload—a failure a confident completion report and passing build could miss. They show how repeatable checks, retained evidence tied to the current code and clear failure guidance can give agents a basis to fix problems before handing work to a human.
The Highest Loop Is People Creating the Next Loop
In The Highest Loop, Shawn Wang argues that AI engineers need to move between levels of work, from debugging tokens and tool use to improving larger goals and automations. He extends that stack to human organization, from small teams to wider communities, and casts AI Engineer World’s Fair as a place where builders can decide what processes to create next. For Wang, the highest loop is the one that brings people together to make new loops.
Odysseus Separates the Agent Harness From the Model It Runs
PewDiePie’s Odysseus separates the agent application from the model it uses: the workspace coordinates tools and information, while Ajax, a fine-tuned Qwen 3.5 9B model, is intended to make tool-use decisions. In his breakdown, Alejandro AO argues that building a useful local agent depends not just on running software on your own machine, but on where the model runs and whether it can reliably use the harness’s tools. He explains how the agent loop works and why training on traces from the target setup may help a smaller model perform better.
Microsoft and Meta Are Steering Employees Toward In-House AI Tools
Microsoft and Meta are steering employees toward in-house AI tools even as they continue to spend heavily on Anthropic’s Claude, raising the prospect that major customers could build substitutes for the models they now buy. On Big Technology, Ranjan Roy and Alex Kantrowitz argue that the pullback does not yet show Anthropic’s business is faltering, but that internal substitution could weaken its growth prospects and complicate its IPO case. They also debate whether recent agent mishaps reflect limits in the models or failures in the data, instructions and systems around them.
Private Document Processing Requires Controlling the Agent’s File Access
Filip Makraduli of Superlinked argues that sensitive document processing should happen before a coding agent sees the file: open models running on infrastructure the user controls can handle OCR, extraction and redaction, then return a smaller artifact for the agent to use. His Obsidian and MCP workflow is designed to reduce exposure to external model APIs and avoid spending context on raw documents, while letting teams choose or adapt models for specific tasks. But a private processing path does not alone guarantee privacy if the agent can read the original file first.
AI Gateways Let Teams Change Models Without Rebuilding Their Stack
As AI budgets face closer scrutiny, Remy Guercio of Tailscale argues that standardizing on one vendor is not a lasting answer to rising costs or AI lock-in. In his account, teams need room to change models, connectors, interfaces and execution environments without replacing the whole stack. An AI gateway, he says, can preserve that choice while giving organizations a central way to manage access, set model-specific budgets and inspect what agents do.
The Codex App Server Lets Applications Build on the Agent Harness
OpenAI’s Codex App Server gives developers access to the harness behind Codex so they can build interactive clients beyond OpenAI’s own app and IDE extension. Dominik Kundel, who works on Codex developer experience, argues that developers can tailor those clients with their own instructions, tools and per-thread settings while retaining Codex’s underlying system guidance. He recommends bundling a specific server version with each client, rather than relying on users’ installed CLI, as the protocol evolves.
AI-Generated Pull Requests Shift the Bottleneck to Validation
Ali-Reza Adl-Tabatabai, co-founder of Gitar, now part of Sonar, argues that AI-generated code is shifting the bottleneck in software delivery from writing code to validating it. More and larger pull requests put pressure on CI and code review, leaving teams to slow delivery or risk rubber-stamping defects, he says. His proposed response is agentic validation: software that reviews pull requests, diagnoses and fixes CI failures, and can approve and merge changes under rules set by the team.
ChatGPT Extensions Keep Product Interfaces and Context Beside the Conversation
Corey Ching argues that ChatGPT plugin extensions can keep a product’s interface and selected context beside a conversation, so users can ask for help without losing sight of what they are working on. The Figma and Meetings demonstrations show how selected designs or meeting notes can inform a request while remaining available for review. Ching recommends starting with a user task, then building and refining an extension around it.
Persistent Agents Turn AI Work Into Shared Team Investigations
At OpenAI DevDay 2026, Jason Liu argues that the next step for workplace agents is not simply helping one person do more, but letting teams contribute to work that continues across conversations, meetings and devices. He says persistent agents, shared workspaces and reusable tools can keep an investigation moving without requiring one person to gather every clue, relay updates and carry the task to completion.