--- title: "Foundations" canonical: "https://current.tinyfish.ai/issue/27/foundations" publication: "TinyFish" publisher: "TinyFish" language: "en-US" guide: "https://current.tinyfish.ai/llms.txt" permalink: "https://current.tinyfish.ai/issue/27/foundations" issue: 27 section: "foundations" description: "Conceptual clarity earned from building at scale" authors: ["Nora Kaplan","Rina Takahashi"] published: 2026-05-04 updated: 2026-04-30 articles: 18 --- # TinyFish > CURRENT is a journal about AI agents in production, for the engineering and product leaders who have to make them work. Drawing on experience building enterprise web agents, it explains why the demo succeeded and the deployment didn't, and what that means for choosing vendors, briefing the board, and avoiding mistakes others have already made. [Read on the website]() ## Issue 27 — Foundations Conceptual clarity earned from building at scale ### The Governance Gap - [The Routing Decision Nobody Governs]() — By [Nora Kaplan]() · Published 2026-04-29 · Format: The Governance Gap Every major agent tracing framework records four identity attributes: description, ID, name, version. None of them include which model is actually reasoning. In production architectures where a planning step runs on one model and execution routes to another, the trace can tell you what happened. It cannot tell you who decided. ### Routing In Practice - [Who Decides Which Model Handles What]() — By [Rina Takahashi]() · Published 2026-04-29 · Format: Routing In Practice When a multi-model agent workflow fires off a request, something picks which model handles it. A developer binding a provider in code. A gateway applying fallback rules. A default nobody remembers setting. Most teams treat this as plumbing and move on. ### Further Reading - **Stanford HAI 2026 AI Index: Technical Performance** — [Source]() — Published 2026-04-29 The primary-source benchmark data showing why no single model dominates across tasks, and why the gaps between capabilities are widening, not narrowing. - **New LLM Releases: April 2026** — [Source]() — Published 2026-04-29 The most complete record of what actually shipped this month. The 50x price spread within a single provider's catalog tells you more about the multi-model future than any strategy deck. - **2026 MCP Roadmap** — [Source]() — Published 2026-04-29 Auth, audit trails, gateway behavior: MCP's own roadmap names the enterprise gaps that compound every time you route a request through a different model. - **Work Moved Into the Browser. Security Didn't Follow.** — [Source]() — Published 2026-04-29 The core insight here applies well beyond browsers: when attribution no longer implies agency, session identity in multi-model pipelines becomes structurally ambiguous. - **Best Multi-Agent Frameworks in 2026** — [Source]() — Published 2026-04-29 A useful survey of which frameworks actually let you swap models at the agent level versus which ones quietly lock you to a single provider's SDK. - **Agent Observability 2026: Evals, Traces, and Cost Attribution** — [Source]() — Published 2026-04-29 Traces fragment at model boundaries. This piece compares the tooling options for cost and latency analysis when a single request touches multiple providers. ### Model Selection Map April 2026 has real competition at every layer of the model stack. Coding, context length, multilingual coverage, cost, local speed. And crucially, the leaders in each category are different models. A top coding score tells you almost nothing about context handling or price efficiency. We organized this month's landscape by what you're trying to build, not who released what. Six use cases, each with the models pulling ahead and the pricing signals that actually matter for agent workloads. These are decision cards, not rankings. In a landscape this dense, knowing which model fits which job beats picking a favorite. - **Claude Opus 4.7 Sets the Coding Ceiling** — [Source]() — Published 2026-04-29 Tag: Autonomous Coding 87.6% on SWE-bench Verified, up from 80.8% on its predecessor. At $5/$25 per million tokens, it's priced as a premium specialist. One wrinkle worth budgeting for: a tokenizer change inflates input token counts by up to 35%. - **Ten Million Tokens Rewrites the Retrieval Math** — [Source]() — Published 2026-04-29 Tag: Large Context Llama 4 Scout offers a 10M-token window starting at $0.08 per million input tokens. Gemini 2.5 Pro hits 91.5% on long-context comprehension at 128K. Both shrink the need for chunked retrieval, though "lost in the middle" problems grow with the window. - **Qwen 3 Stretches to 119 Languages Natively** — [Source]() — Published 2026-04-29 Tag: Multilingual Agents Expanded from 29 to 119 languages (Qwen 3.6 reaches 201). Its thinking-budget control lets agents vary reasoning depth per subtask, handy when deep translation precedes structured extraction. Open-weighted under Apache 2.0, so you can actually inspect what you're deploying. - **Mistral Medium 3 Makes Agent Loops Affordable** — [Source]() — Published 2026-04-29 Tag: Cost Optimized $0.40 input, $2.00 output per million tokens. Roughly a quarter of what Opus 4.7 costs. Mistral claims performance within 90% of Sonnet 3.7, though independent verification is still catching up. Treat that number as directional, not settled. - **Local Models Cross the Agent-Ready Latency Threshold** — [Source]() — Published 2026-04-29 Tag: Local Inference Ollama crossed 52 million monthly downloads in Q1 2026. Sparse models like Qwen3-Coder-Next activate 3B of 80B parameters, hitting 40–60 tokens/second on a 16GB laptop. Gemma 4 adds native function calling. The practical play is hybrid: local for speed, API for ceiling. - **Strength in One Category Predicts Nothing Elsewhere** — [Source]() — Published 2026-04-29 Tag: Jagged Frontier Opus 4.7 scores 87.6% on SWE-bench Verified but drops to 64.3% on SWE-bench Pro. Gemini leads long-context benchmarks and trails on coding. April 2026 has enough model density to route across specialists. The landscape rewards orchestration over allegiance. ### Routing & Observability Assigning different models to different agents is now a one-liner in every major framework. That capability arrived fast and arrived everywhere. What didn't arrive with it: the ability to actually govern what happens next. Trace which model made which decision. Attribute cost across providers in a single run. Understand what the orchestration layer did with the routing logic you handed it. Across four platforms, a pattern emerges. Configuration is generous. Visibility is improving. Control over what sits between them remains surprisingly thin. - **Microsoft's Six Providers Hide Behind One Preview Tool** — [Source]() — Published 2026-04-29 Tag: Model Routing Version 1.0 connects Azure OpenAI, OpenAI, Anthropic, Bedrock, Gemini, and Ollama. Mixing providers across agents in a single workflow takes one line of config. Seeing what actually happened requires DevUI, still in preview. Full cost attribution needs Azure infrastructure underneath. The on-ramp is wide open; the rearview mirror is fogged. - **LangGraph Locks Model Choices Into Compiled Paths** — [Source]() — Published 2026-04-29 Tag: Graph Orchestration Any Python function becomes a node, any provider can power it, and conditional edges route execution based on state. Once compiled, though, the graph is frozen. LangSmith traces surface provider metadata per run, which is genuinely useful. Cross-model cost attribution? That's your problem. Manual instrumentation, operator by operator. - **CrewAI Exposes Model Choice, Keeps Delegation Opaque** — [Source]() — Published 2026-04-29 Tag: Agent Assignment Each agent takes an explicit llm parameter. Hierarchical crews get a separate manager_llm setting. Clean so far. But how that manager actually delegates work is framework-internal, not operator-configurable. Cross-model tracing in open-source requires bolting on separate instrumentation per provider. Unified visibility lives behind the Enterprise paywall. - **Google Shows Everything, Lets You Configure Less** — [Source]() — Published 2026-04-29 Tag: Platform Observability The Gemini Enterprise Agent Platform ships the richest native observability of the group: DAG traces, topology graphs, OTel-compliant telemetry, live evaluation scoring, and access to 200+ models via Model Garden. Impressive instrumentation. Yet Agent Memory Bank documentation is silent on what happens when different models write entries to the same identity-scoped store. The platform can tell you what occurred. Whether the memory is coherent is a question it doesn't surface. ## Other sections in Issue 27 - [Daily Brief]() - [Market Pulse]() - [Practitioner's Corner]() - [Echoes]() - [Vision]() ## More - [Publication guide]() - [Issue archive]()