Field Notes

Make agents secure and scale them: PKI for MCP, runtime discovery, and a voice SDK

You build agents that access internal systems. Today is about identity, runtime integration, and fast day-to-day rollouts.

Key AI news

AWS + Keyfactor: Cryptographic identity for agents

AWS and Keyfactor show a reference architecture that gives agents unique X.509 certificates. It signs every request, links it to a user signature, and checks revocations in real time via OCSP. This closes a gap in the MCP ecosystem. MCP itself defines no authentication. In the example the agents run on Amazon Bedrock and the Strands Agents SDK. The code is ready. If you use agents in sensitive workflows, you need separate user and agent certificates with short lifetimes and central revocation. This is available, but it is a reference, not a managed service.1

AWS Agent Registry: Find tools at runtime instead of hardwiring them

The AWS Agent Registry catalogs MCP servers, agents, and skills. An orchestrator searches at runtime for semantically fitting capabilities, instantiates clients, and executes. The registry does not provide credentials by design. You fetch those separately. This decouples team releases and saves redeploys once a new tool is registered. The example uses Bedrock AgentCore, Cognito, and Gateway. This is available, with example code.2

Breaking: Eight eMudhra intermediates compromised

The private keys of several eMudhra intermediate CAs have leaked. They chain to widely used root stores. Update your CA bundles on servers, laptops, containers, and CI images. Python certifi and common package sources have shipped updates. This also affects internal agents that use TLS. The notice is from today.3

Mistral Voxtral 3: Voice agent SDK for 50+ languages

Mistral presents a voice SDK with real-time recognition, emotional TTS, and deep tool use. You can host it in the EU or self-host it. It targets service hotlines, voice assistants, and internal voice UIs. Prices are not in the post.4

OpenAI acquires Orchestra Bio

OpenAI buys Orchestra Bio to speed life sciences workflows with AI. This is an announcement without an immediate product. Relevant for biopharma teams that hope for new tools.5

Highlights for your daily work

  • Cash reconciliation in hours instead of days. A US franchisee reports 94 percent less time for cash reconciliation and 95 percent faster menu repricing rounds with ChatGPT Work. Numbers are from the vendor. If you have manual reconciliations, define a small sub-process and have checklists, table templates, and variance lists generated.6
  • Keep context in agents clean. DeepLearning.AI offers a free course on context pruning, compaction, scratchpads, and agentic search. Start with clear rules for what gets dropped from history after each task and what stays as a reusable note.7
  • Better decisions with rubrics. Terminal X reports 88.6 percent agreement with senior investors on assessment, plus 25 percent more rule violations found. Numbers are from the vendor. Write a 10-point rubric for your reviews and have drafts scored automatically up front.8

Tools and updates

  • Linear attention for long contexts. i14s Laplace 3B keeps the inference state per sequence at a fixed 29.5 MB. On a trn2.48xlarge it trained at about 120,000 tokens per second and served 11,544 tokens per second across 1,280 parallel sequences. Good for always-on agents and very long documents. Requirement: AWS Trainium2 and TorchNeuron Native (beta). You must validate recall quality per workload.9
  • Local AI on the laptop. NVIDIA RTX Spark is available and combines a Blackwell RTX GPU and a Grace CPU with up to 128 GB of shared memory. Starting price 2,399 US dollars. Useful for confidential or latency-critical workflows without cloud.10
  • More stable agent orchestration with n8n 2.45.0. Fixes include clearing inactive MCP browser sessions, deleting agent caches after credential updates, and gateway model IDs. If you build agents with n8n, these changes raise reliability.11
  • External review of safety tests. Amazon had AVERI review its own safety evaluations. Result: expanded multilingual coverage, introduced difficulty tiers, quarterly updates. Apply the idea of an independent meta-evaluation to your internal tests.12

Try it in five minutes

1) Create a short scoring rubric. Pick a frequent output, like a product description. Define 8 to 10 criteria with a 1–5 scale and short notes.8 2) Let it score. Paste rubric and two sample texts into your existing AI chat and ask for structured JSON scores plus justifications per criterion.8 3) Cross-check. Compare the three most important criteria with a human rating. Expected result: clear agreement on core errors. Checkpoint: at least 70 percent overlap on the top 3 criteria.8

A good find

One Useful Thing looks back on 30 months of AI progress. Three points carry into daily work: You work less in chat and more as a conjurer who starts agents on cue, lets them generate variants, and curates results. The author says that on the METR benchmark the capabilities doubled eight times until the measurement broke. Cost per hard answer fell from 30 cents to 0.0004 US cents. These are vendor and author claims without independent verification. Consequence for you: delegate end-to-end tasks to an orchestrator instance, ask for options instead of a single output, and have the AI mark its reasoning so you can check it.13

Sources

  1. Securing AI agent identity with X.509 PKI on AWS (aws.amazon.com)
  2. Discover tools at runtime with AWS Agent Registry (aws.amazon.com)
  3. eMudhra Intermediate CA Compromise (simonwillison.net)
  4. Voxtral 3 (mistral.ai)
  5. OpenAI acquires Orchestra Bio (openai.com)
  6. Franchisee reconciles cash in 94% less time with ChatGPT Work (openai.com)
  7. Learn to build to better AI Agents through better Context (youtube.com)
  8. Terminal X matches senior investor judgment 88.6% with GPT-6.1 (openai.com)
  9. Beyond the KV cache: Training and serving i14’s Laplace 3B linear attention model on AWS Trainium2 with native PyTorch (aws.amazon.com)
  10. RTX Spark Ignites Global Celebration as New Windows PCs Arrive (blogs.nvidia.com)
  11. n8n@2.45.0 (github.com)
  12. Strengthening AI safety evaluation through independent third-party review (amazon.science)
  13. 30 Months on the Exponential (in Charts) (oneusefulthing.org)