--- title: "Stanford gives AI slop an F" newsletter: "User Community" date: 2026-11-05 source: https://aaif.live/newsletters/usercommunity/2026-11-05-stanford-gives-ai-slop-an-f --- # Stanford gives AI slop an F *Plus… agents with credit cards, Git at Codex scale, and why TOON increased token costs.* *User Community — Agentic AI Foundation, 2026-11-05* Expect AI agent progress to stall now that they’ve invented middle management [https://the-decoder.com/one-brutal-minecraft-night-scared-50-ai-agents-into-inventing-chicken-handoff-protocols-a-fertilizer-officer-and-fishing-shifts/]. ## Gems ## Scaling discovery requires forgetting ## AAIF Reading Group CAN FORGETTING HELP AI AGENTS FIND BETTER SOLUTIONS? More context doesn't always mean better results. The paper Scaling Discovery Requires Forgetting [https://www.alphaxiv.org/abs/2610.20563] explores how periodically resetting an agent's history, while keeping its best solution, can improve performance. Join Sparsh Jain and Arthur Coleman to explore why agents hit a plateau, when forgetting helps, and what this means for building better autonomous research agents. Thursday, November 12 | 9 AM PT / 12 PM ET / 5 PM GMT / 6 PM CET Register for the free virtual reading group. [https://home.mlops.community/home/events/aaif-reading-group-scaling-discovery-requires-forgetting-pg1zmi4qzx?agenda_day] ## What’s happening across the chapters ## LOCAL ORGANIZERS It was great to see some well-deserved recognition at AGNTCon + MCPCon North America, with AAIF ambassadors and community organizers celebrated on stage. Special congratulations to Agentic Tokyo, which received the Community Builder Award at the inaugural AAIF Annual Technical Excellence Awards! Satoshi Ito shared the news [https://lnkd.in/p/e4KQwSm8], while Vijay Bhore [https://lnkd.in/p/eVtUTzRN] and Yelaine Wang [https://lnkd.in/p/etiSt6BZ] also reflected on the recognition and the people behind the community. It’s been a busy week across our local chapters, with seven meetups in New Zealand, Luxembourg, Denver, Pune, Shenzhen, Chennai, and Aalborg. Speakers covered everything from hands-on agent building and MCP to production architectures, security, and even live robotics demos. Catch up on all seven in AAIF’s community roundup [https://www.linkedin.com/posts/agentic-ai-foundation_seven-cities-seven-meetups-one-growing-activity-7520903121989079041-kzLO], and thanks to the organizers, speakers and supporters who made them happen. ## How Stanford teaches AI coding now In one Stanford course, students submit open-source contributions as homework, and a maintainer complaint about slop means a zero. Their lecturer rewrote most of the course after Opus 4.5 and also builds agents for a sales platform. * Students research and plan before implementing, and agents check their work with Playwright through MCP or end-to-end tests. * Rather than mapping tools directly to API endpoints, his team built a single meetings query tool supporting keyword and semantic search. * His team compiles natural-language agent descriptions into a structured intermediate representation, then into LangGraph, so they can swap workflow engines. Engineers can improve agent workflows by gathering context first and designing tools around tasks. [https://podcasts.apple.com/gb/podcast/how-stanford-teaches-ai-coding-now/id1505372978?i=1000799479080](https://podcasts.apple.com/gb/podcast/how-stanford-teaches-ai-coding-now/id1505372978?i=1000799479080) [https://home.mlops.community/home/videos/how-stanford-teaches-ai-coding-now](https://home.mlops.community/home/videos/how-stanford-teaches-ai-coding-now) [https://open.spotify.com/episode/6abjF5dcxkrbvupxycPnpk?si=Dst28BeUQx22NfwT3FwJpA](https://open.spotify.com/episode/6abjF5dcxkrbvupxycPnpk?si=Dst28BeUQx22NfwT3FwJpA) ## Building AI so everyone can flourish An AI research assistant recommended hiring another engineer; after consulting the Google SRE handbook, it instead called for an operational toil audit. That reversal frames a wider question: how should human expertise shape increasingly autonomous AI systems? * Context engineering: Curated expert knowledge can redirect an LLM from surface symptoms to underlying organizational problems. * Agent infrastructure: Inconsistent shell environments and tool configurations trigger repeated failures and wasted tokens; shared setups can preserve hard-won fixes. * Open architectures: MCP, portable memory, and interoperable protocols offer ways for agents and people to cooperate across platforms. Reliable AI depends on deliberately choosing the knowledge, constraints, and incentives guiding its work. [https://podcasts.apple.com/gb/podcast/tim-oreilly/id1505372978?i=1000798194214](https://podcasts.apple.com/gb/podcast/tim-oreilly/id1505372978?i=1000798194214) [https://home.mlops.community/home/videos/building-ai-so-everyone-can-flourish](https://home.mlops.community/home/videos/building-ai-so-everyone-can-flourish) [https://open.spotify.com/episode/2uR1Rbz4Gw0gFYefXqIDZw?si=gXa0p2U4RSqX5XYjH1RcIA](https://open.spotify.com/episode/2uR1Rbz4Gw0gFYefXqIDZw?si=gXa0p2U4RSqX5XYjH1RcIA) ## From slop code to software factories Coding models diagnose bugs well, but months of unread AI-generated code can fail in ways they can't fix. Riding a tandem bike through Amsterdam, two builders discuss harness design and agent infrastructure. * Codex edits better with its own patch tool than with Claude Code's string replacements, so tool schemas should match model training. * Opus 4.8 turned two hours of footage into contact sheets and a storyboard, then refined the demo through ten queued prompts. * Teams should own dev environments but reuse infrastructure for streaming agent output to Slack and sharing diffs before pull requests. Builders can adapt tools to model behavior without rebuilding collaborative infrastructure. [https://podcasts.apple.com/gb/podcast/from-slop-code-to-software-factories/id1505372978?i=1000798752135](https://podcasts.apple.com/gb/podcast/from-slop-code-to-software-factories/id1505372978?i=1000798752135) [https://home.mlops.community/home/videos/from-slop-code-to-software-factories](https://home.mlops.community/home/videos/from-slop-code-to-software-factories) [https://open.spotify.com/episode/2s43sNyIvozm7fVEGYGrEY?si=16X5yurQSfqWfQURGfWATg](https://open.spotify.com/episode/2s43sNyIvozm7fVEGYGrEY?si=16X5yurQSfqWfQURGfWATg) ## Your AI can click “buy.” Should it? When an attacker hides extra payment instructions in an invoice, a finance agent faces the confused deputy problem from 1988. Two practitioners argue that agent authorization depends on knowing what task each action serves. * They borrow fraud detection's layered approach, running hard-coded policies before costly LLM judges, which attackers can fool with plausible text. * Each agent needs its own workload identity under a human principal, and email access should expire once the agent finishes reading. * Computer-use agents inherit a browser's saved sessions and cards, so anyone who reaches that machine can act as the user. Teams can contain injected instructions and overprivileged agents by checking each action against the task it serves. [Video](https://home.mlops.community/home/videos/your-ai-can-click-buy-should-it) ## Please stop telling your AI to “search harder” Kubernetes was designed to manage thousands of services, but platforms such as Lovable could need to run millions of distinct agent sessions. Two practitioners explore what bursty agent traffic means for systems built around predictable machine users. * Agents can fire tens of searches per question, so multi-year API gateway contracts priced per request are harder to forecast. * In Agent Substrate, an early Google project built on Kubernetes, mostly idle agents share a pool of workers and resume in under a second. * They expect most teams to use Substrate through a managed platform, since few companies run agents at consumer scale. Planning for agents at scale means accounting for both unpredictable API traffic and stateful sessions that need to stop and restart quickly. [Video](https://home.mlops.community/home/videos/please-stop-telling-your-ai-to-search-harder) ## Where voice agents fail in production: Part 2 - tools, recovery, and observability A voice agent can say “Booked” even when the reservation failed, changed, or happened twice. Preventing that mismatch requires more than accurate speech recognition. * When a caller interrupts, external actions may keep running, so teams should use stable operation IDs, idempotency, and system-of-record checks before retrying timeouts. * On 278 grounded tasks, text-based GPT-5 reasoning reached 85% completion, while tested voice agents managed 31–51% in clean conditions. * Teams can debug recovery by tracking handoff acceptance and pending actions, then correlating audio, playback, tool results, and state transitions with shared IDs. Reliability depends on verifying what the system actually did, not what the conversation seemed to resolve. [Read the blog](https://aaif.io/blog/where-voice-agents-fail-in-production-part-2-tools-recovery-and-observability) ## We added TOON to our MCP server, then removed it When a web-scraping team shipped TOON, a JSON alternative whose README claimed 30–60% fewer tokens, their busiest MCP tool got more expensive. They pulled it after 30 days and two A/B rounds. * The four list tools that shrank 19–44% under TOON received only 0.07% of all calls. * Nested scraped records don’t fit TOON’s table layout, so keys repeat. On 100 Google Maps places, tokens rose 11.7% over compact JSON. * When agents requested three fields instead of about 60, a 1 MB payload fell to 12 KB. TOON reduced the same payload by 14% in bytes. Teams can judge other context optimizations the same way, by counting tokens per tool and weighting by call volume. [Read the blog](https://aaif.io/blog/we-added-toon-to-our-mcp-server-then-removed-it) ## Total cost of AI over value realized In one enterprise audit, direct model consumption accounted for roughly a quarter of total AI spending. That leaves teams measuring more than token bills when assessing whether agents pay off. * Teams should count infrastructure, routing, observability, retries, idle compute, and human approvals in the cost of completing each task. * A cheaper model can cost more if its errors require supervision or hurt retention, so only acceptable outputs count toward value. * To show realized value, teams should distinguish capacity gains, augmentation, and automation, then connect claims to revenue, removed or avoided spending, or freed capital. Comparing Total Cost of AI with verified Value Realized makes the trade-off between spending and outcomes visible. [Read the blog](https://aaif.io/blog/total-cost-of-ai-over-value-realized) ## In-person events ## Come and connect * Melbourne [https://luma.com/7qtmasup] - November 5 * Helsinki [https://luma.com/rzrt9958] - November 5 * San Francisco, Voice Agents Forum [https://luma.com/voiceagentssf] - November 5 * Mumbai [https://luma.com/sywq28p8] - November 7 * London, Workshop [https://luma.com/uiowlg0f] - November 10 * Ulaanbaatar [https://luma.com/vczb32ch] - November 15 * Pune [https://luma.com/gb5bwpen] - November 28 Find your city here [https://aaif.io/events?tab=community], or start a chapter if there isn't one yet. ## Virtual events ## Join from anywhere * Lunch and Learn: Exploring the Future of AI-Assisted Software Development [https://home.mlops.community/home/events/coding-agents-lunch-and-learn-session-30-exploring-the-future-of-ai-assisted-software-development-7z5yrenoja?agenda_day=] - November 6 * AAIF Reading Group: Scaling Discovery Requires Forgetting [https://home.mlops.community/home/events/aaif-reading-group-scaling-discovery-requires-forgetting-pg1zmi4qzx?agenda_day] - November 12 --- Source: https://aaif.live/newsletters/usercommunity/2026-11-05-stanford-gives-ai-slop-an-f