Field Notes

Build agents that fail less: context under control, smart approvals, better images

Today you get concrete steps to run agents cheaper, ease approvals, and ship visual assets faster.

Important AI news

Four layers to cut context bloat in production agents

A new practical post maps context engineering techniques. That means targeted optimization of the model context. It puts them into four layers and shows code. Architecture, prompt rules, file offloading, and independent checking agents mesh together. In the reference build, extra tokens per script with a write-and-execute tool drop by about 95 percent. The system prompt shrinks by around 97 percent with on-demand skills. These are vendor numbers from the shown Deep Insight solution, built on Strands Agents SDK and Amazon Bedrock. If you run agents for data analysis or reports, you can cut costs and errors noticeably.1.

Kimi K3.1 Pro: open-weight multimodal model with 1M context

Moonshot publishes Kimi K3.1 Pro on Hugging Face. It is an open-weight model. The weights are public. The license is “other” and adds usage limits. It offers 1,048,576 token context and uses an MoE design with 2.8T parameters, 104B active. The many benchmarks are vendor measurements and partly internal. Check if they fit your use.2.

OpenAI Images 2.6 launches with faster generation and GIFs

OpenAI introduces Images 2.6. New: faster images, GIF creation, more precise edits, and extra API controls. The vendor says this applies to ChatGPT, Codex, and the API. For you that means: You produce assets and small animations directly in your existing workflow.3.

OpenAI acquires Pierre Computer Company

OpenAI announces the acquisition of Pierre Computer Company. The goal is to expand Codex tools for building, understanding, and checking code. This is an announcement. There are no concrete new features for you yet.4.

Turn agent failures into tests

NVIDIA shows how to turn production evidence into prioritized issues and domain-specific evals. That way you improve prompts, tools, models, memory, and harness systematically. Useful when generic benchmarks do not explain your errors.5.

Highlights for your workday

  • Smart Approvals in Amazon Quick cuts needless back-and-forth without disabling approval duties. A separate classifier checks proposed actions against your task and context. You can define what always goes to a person and see evidence in CloudTrail. Next, review connector permissions for sensitive actions.6.

  • Documents as structured fields: The Dharma Field Extractor takes arbitrary files and returns JSON. This helps with invoices, resumes, or forms when you must transfer fields into systems. Start with a small set of sample files and compare extracted values to your required fields.7.

  • Kino AI brings a collaborative, code-first video editor with agent support. The full editor is free. Pro costs 25 US dollars per editor per month. You need the Electron app. Good for teams that iterate social or product videos regularly and want agents to assist.8.

Tools and updates

  • Datasette Apps 0.3a0: LLMs can reference apps in Datasette more easily and manage permissions for saved queries. Write access requires your confirmation. Requires an existing Datasette instance, alpha status.9.
  • Together Compiler: Optimizes open models within 24 hours with agents and persistent engines. Up to 3.7x more throughput per vendor. For teams with their own inference ops, technically demanding.10.
  • Polymorphic Inference: Self-tuning for open models across changing loads and chips. Works with Together Compiler. Mostly relevant for operators with variable demand.11.

Try it in five minutes

Create a short product GIF in ChatGPT with Images 2.6

1) Open ChatGPT and enter: “Create a three-part, seamless looping 512x512 GIF for a ‘New feature’ announcement. Colors: #0A84FF, #111827, white. An icon appears, turns into a checklist, then into a checkmark.”3. 2) Ask for a variant with a precise edit: “Swap the checklist for a gear. Set frame duration to 0.7 s per frame. Sharpen edges for web display.”3. 3) Download the GIF and check it at 200 px size.

Expected result: A clearly readable, smoothly looping GIF in your colors. Checkpoint: File size under 1 MB, elements are recognizable at 200 px without pixel mush.3.

A good find

Practical AI distills three pragmatic points. Build reliable workflows with suitable guardrails and evals. Co-develop with the people on the front line, not just for them, and tune autonomy to goals and review paths. Take one thing from it. Define a clear human checkpoint for one agent today and measure its effect. Based on the show notes.12.

Sources

  1. Four layers of context engineering for production AI agents: How Deep Insight applies them in working code (aws.amazon.com)
  2. moonshotai releases moonshotai/Kimi-K3.1-Pro on Hugging Face (huggingface.co)
  3. Introducing Images 2.6 (openai.com)
  4. OpenAI to acquire Pierre Computer Company, Inc. (openai.com)
  5. Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes (developer.nvidia.com)
  6. Smart Approvals in Amazon Quick: Human oversight where it matters (aws.amazon.com)
  7. Dharma Field Extractor: Any Document In, Structured Fields in JSON Out (huggingface.co)
  8. Launch HN: Kino AI (YC S23) – AI media manager and video editor (news.ycombinator.com)
  9. datasette-apps 0.3a0 (simonwillison.net)
  10. Together Compiler: Continuous Compilation for AI inference (together.ai)
  11. Polymorphic Inference: self-tuning inference for open models (together.ai)
  12. Abundant Intelligence at the Midwest AI Summit (share.transistor.fm)