This post consolidates what OpenAI, Anthropic, Google and the MCP spec recommend for designing tools for AI agents, and covers what their guides leave out. It argues tools should be built around workflows rather than API endpoints, that names, descriptions and schemas act as a contract the model reads literally, and that responses and errors are part of the prompt.
Stanford's CS 329Z, Engineering AI Agents, walks through the anatomy of agentic systems — LLM core, reasoning, memory, tools and environment — and the gap between compound AI systems and true agents. The course covers RAG pipelines, tool use and function calling with MCP, DSPy and LangGraph orchestration, memory architectures such as MemGPT, and multi-agent orchestration patterns. It closes with optimization methods including GEPA, PPO and GRPO, plus a guest lecture from the Terminal-Bench creators.
The AI Security Wiki is a reference for securing enterprise AI adoption, covering governance, risk management, secure engineering, operational monitoring, vendor evaluation and assurance in reusable template pages. It maps control objectives, audit evidence and test cases onto frameworks including ISO/IEC 42001, NIST AI RMF, NIST SSDF, PCI DSS v4.0.1, the FINOS AI Governance Framework and CERT-In's AI security auditing guidelines. It also includes a 682-case test library drawn from OWASP lists, MITRE ATLAS and NIST AI 100-2.
Solibra built Sovereign, a router that sits between an application and model providers and learns per-request model choice from grades of its own finished work, with no answer key and no prior knowledge of the models. Across eight public benchmarks run against Claude Opus 5.5 and Claude Haiku 5.5, Sovereign spent 43% to 81% less than Opus on every benchmark, matched or beat Opus accuracy on USACO, LiveCodeBench and SATBench, and was more accurate than Haiku everywhere, by up to 55 points on USACO. On USACO its accuracy rose from 88% over the first 25 tasks to 97.7% by the end, passing Opus around task 150, and its grader agreed with the hidden answer key 95% of the time.
ElevenLabs outlines seven customer service skills — communication, problem-solving, conflict resolution, empathy, adaptability, personalization and product knowledge — that both human and AI support agents need. It cites a Zendesk study finding over 50% of customers switched to competitors after one bad experience, and a Gartner forecast that at least 70% of customers will start their journey with an AI interface by 2028. The guide stresses evaluating conversational metrics during training and after deployment, plus seamless handoff to humans for de-escalation.
Databricks details how credential-driven attacks unfold on its platform and how verbose audit logs plus Genie can detect the behavioral shift after a stolen PAT, session cookie, or service-principal credential is used. The post maps adversary activity into six stages, from initial login through environment reconnaissance, and covers response steps and controls that limit what a stolen credential can reach.
Bright Machines now gives customers live, governed dashboards of factory data from air-gapped plants without opening any inbound access. It captures PostgreSQL changes at the edge with Debezium, writes them as Apache Iceberg, pushes them out over one auditable outbound path, and reads them in place in Databricks via Unity Catalog federation. The first customer-visible dashboard went live about two months after kickoff, with committed rows queryable in one to two minutes and dashboards refreshing every ten minutes.
OpenRouter benchmarked 11 rerankers on a 274-question synthetic RAG dataset and found that reranking the top 30 vector-search candidates with voyageai/rerank-3-lite raised precision@1 from 81.0% to 97.1% at about $0.10 per 1,000 queries. The guide recommends retrieving roughly 30 chunks, reranking them, and sending only the top 3 to the LLM, which in its test lifted answer accuracy to 99.3% while cutting answering cost from $0.22 to $0.16 per 1,000 questions. It also flags failure modes: omitting the input_type parameter cost 7 to 12 points of precision@1, and cohere/rerank-v3.5 silently ignores content beyond its 4,096-token limit.
A developer reverse-engineered the Wi-Fi-enabled Coway Airmega IconS air purifier to cut its dependence on Coway's IoCare cloud, using a man-in-the-middle setup built with LLM assistance. The existing Home Assistant integration impersonates Coway's phone app but breaks repeatedly—server bans, an IoCare to IoCare+ rebrand, HTML layout changes, and outages—prompting the effort. A weak-key incident at certificate authority eMudhra in October 2026 opened a possible foothold for intercepting the device's traffic.
GitHub published a practical guide to agent skills, versioned directories centered on a SKILL.md file that describe when and how to perform a repeatable task. The guide uses two examples, Proof Agent for adversarial verification of AI-generated code and Data Forge for a multi-skill data pipeline, to show how skills share a project manifest, reuse one execution path, and place human approval at each phase. It also describes a packaging script that generates client-specific files for Claude Code, Cursor and GitHub Copilot from a single skills source.
Natera built SWARM, an agentic AI system for lab operations, on Amazon Bedrock AgentCore, cutting first response time by about 130x and automating 96.2 percent of support tickets. The system pairs a swarm-agent that drafts diagnostic reports before engineers open a ticket with a swarm-hive for real-time engineer interaction, both sharing one skill library. It addresses over 65 daily Jira tickets across three queues that previously averaged 17.4 hours for first response.
Databricks responds to four common misreadings of LTAP, the Lakebase architecture that keeps one governed copy of operational data serving both Postgres and analytics without CDC. The post argues LTAP is not replication: layers in front of the lake are caches, not a second database with its own catalog. It also says heap pages and Parquet are two lossless encodings of the same rows, so point lookups and scans each read the layout fast for them.
AWS shows how to track and cap Amazon Bedrock generative AI spend per IAM principal using only native features: Bedrock automatically records the IAM user or role on every inference request, and those tags become cost allocation tags readable in Cost Explorer or CUR 2.0 exports. An AWS Budget scoped to Bedrock and filtered to one identity or tag value can attach a budget action that applies a Deny IAM policy once a threshold is crossed, with no application code for the common case.
Amazon Bedrock's cache duration, service tier, reasoning effort, and maxTokens settings can change inference cost without altering the model or prompt. A 5-minute cache write on Claude Sonnet 5 and Claude Opus 5 costs 1.25x the input rate, a 1-hour write costs 2x, and cache reads cost 0.1x; Flex tier requests to unsupported models fail with a ValidationException.
Amplitude's proactive product agent Wave, built primarily on OpenAI models, identifies what to build, helps teams ship changes with coding agents, and measures the impact. The setup lets Amplitude ship product improvements in 1–2 days.
LangChain explains eval-driven development (EDD), a three-step process for building AI agents: define what good looks like, write the eval before you build, and compare results after each change. Because agents are non-deterministic, evals score runs across a dataset of examples and can use LLM-as-a-judge, with real agent runs in LangSmith as a source of test inputs.
Constellation, the largest US nuclear fleet operator, is building a governed Enterprise Data Platform on Databricks to give the business one shared meaning for its data. The platform uses Unity Catalog for lineage, fine-grained access and a single permission model, plus Data Vault modeling for critical reusable entities. Early use cases include a nuclear engineering assistant built with Databricks Genie and outage planning.
Eclipse Aerospace is using Databricks PySpark custom data sources to ingest nearly 20 years of proprietary flight data, over 3TB spanning more than a million flight hours, without always-on dedicated compute. The Rust-backed Python package delivered twice the benchmarked performance of a similar PyArrow-based pure-Python data source and nearly 20x that of a native Python implementation without Arrow. The API, introduced in Apache Spark 4.0 and available in Databricks Runtime 15.4 LTS, replaces the prior Lambda-to-Avro pipeline.
A self-service LLM gateway built on Amazon Bedrock Application Inference Profiles and LiteLLM running on AWS Fargate exposes Bedrock foundation models behind an OpenAI-compatible API, giving developers a single base URL and per-user key. The setup adds per-user spend controls, per-model quota isolation, and CloudWatch RPM/TPM observability, with Langfuse for prompt-level telemetry. Self-service key issuance requires LiteLLM Enterprise, since the free tier is limited to five users.
This AWS post argues that Amazon Bedrock spend should be judged by cost per task rather than cost per token, since tokenizer efficiency, output verbosity, agentic loop count and reasoning overhead can make a lower-priced model cost more. A worked example shows a model with 2x the list price finishing 65% cheaper per task, and the post adds caching, routing and output constraints plus a five-dimension scorecard for measuring agent ROI against human baselines.
Claude Platform engineers Jake Dejno and Lucca Siaudzionis explain how the Claude API evolves alongside each model launch, covering feature release decisions, model versioning and migrations, and cost management. They detail best practices for changing effort mid-conversation without breaking prompt caching, and how Claude Mythos and Fable share the same model weights with different safeguards. Claude Managed Agents and mid-conversation effort changes are in beta, and Mythos access runs through the Cyber Verification Program.
H6.ai automates accounts payable on top of customers' existing ERPs, using a dedicated Lakebase Postgres database per customer as its system of record. Because invoice data keeps changing, the team uses Lakebase copy-on-write branches to freeze snapshots of customer databases and run each regression test on its own short-lived branch, so the live database is never written to. Each run is judged by deterministic assertions plus a panel of AI judges from different model providers.
Amazon Bedrock AgentCore Evaluations now lets teams run AWS built-in evaluators alongside managed DeepEval and AutoEval metrics, plus custom evaluators derived from either. The built-in catalog spans 15 evaluators across response quality, goal completion, tool use, skills and safety, with three trajectory checks for expected tool paths. DeepEval adds multi-turn metrics such as ConversationCompleteness and KnowledgeRetention, while AutoEval contributes specialized signals including security.
The Web Search API in Amazon Bedrock AgentCore is now available, giving agents current web results through a single SigV4-signed API call served from a web index Amazon operates. A walkthrough builds a daily dependency watch as a LangGraph graph that researches npm, PyPI, and container base images in parallel, with date windows, blocklists, and search budgets enforced in code. In testing, the three branches finished in 36 seconds versus 99 seconds run sequentially.
AWS Quick introduces AI-PDLC, a framework of six specialized agents—led by an Orchestrator covering Discovery, Prioritize, Working Backwards, Spec Generator, and Designer—that turns scattered enterprise signals from Slack, Jira, Salesforce and SharePoint into a traceable prototype_spec.md. Every phase requires human approval, and the spec can be handed to coding agents like Kiro or Claude Code via the open Agent Client Protocol (ACP).
AWS has published a catalog of 23 production-ready, open source generative AI assets from its Generative AI Innovation Center, now part of AWS Frontier AI Engineering & Services. The building blocks cover RAG pipelines, document processors, agent scaffolding and evaluation frameworks, and AWS says they can cut development from months to weeks. DoorDash reported a 50x increase in testing capacity and Ricoh a 7-fold throughput increase after building on the assets.
DXC Technology built the model substrate for its DXC OASIS IT-operations harness on Amazon Bedrock, reaching Anthropic Claude, Amazon Nova, Meta Llama and open-weight models through one managed interface. Amazon Bedrock Guardrails apply content filtering, PII redaction and denied-topic policies uniformly across every model, while AWS PrivateLink keeps inference traffic inside the customer VPC. DXC OASIS has been generally available since March 2026 and cut resolution time by 95% on average across 56 measured use cases.
AWS introduces the Agent Development Lifecycle (Agent DLC), a framework with one lifecycle, two loops, and a single explicit standard for judging every agent change. It addresses three agent properties that break conventional CI/CD: non-determinism, prompts as source code, and independently moving dependencies. A reference HR policy assistant shows chunk-level retrieval precision of 1.00 across all three test questions while fact-level precision fell to 0.09 and retrieval coverage hit 0.00 on one question.
AWS shows how to give a shared AI agent per-user document access on Amazon Bedrock Managed Knowledge Base by exposing search as MCP tools through AgentCore Gateway. A request interceptor injects the verified user's email from the Cognito JWT into userContext, and ACL awareness returns only documents that user may view. A Cedar policy in AgentCore and a resource-based policy on the knowledge base add secondary defenses against spoofed or unauthorized searches.
AWS published a reference architecture in which a LangChain-based enterprise orchestrator on Amazon ECS calls a Strands Agents-based compliance SaaS agent on Amazon Bedrock AgentCore over the A2A protocol, with the two sides in separate AWS accounts and no shared secrets. Each side authenticates with tokens from its own Amazon Cognito user pool, and the SaaS agent calls back to an enterprise context agent for account metadata, so the agents act as peers rather than client and server. The system deploys from three Jupyter notebooks that provision five CloudFormation stacks and run an end-to-end test plus seven trust boundary checks.
AWS added a knowledge base migration guide to the amazon-bedrock skill in the Agent Toolkit for AWS, letting a coding agent plan a move from a customer-managed knowledge base to Amazon Bedrock Managed Knowledge Base. The agent starts with read-only discovery and a line-by-line reviewed plan, then creates the new knowledge base and data sources alongside the original or writes a migrated IaC template. Managed Knowledge Base raises default quotas to 200 data sources per knowledge base and 10,000 knowledge bases per account per Region, and adds agentic retrieval, AgentCore Gateway integration, and resource-based policies.
Anthropic's guidance matches each Claude model to the scale of the work: Opus 5.5 for planning and building full features, and Fable 5.2 for the most ambitious, open-ended technical work. Sonnet 5.5 covers everyday tasks such as fixing a bug or drafting a doc, while Haiku 5.5 handles the smaller tasks done again and again.
A developer used Gemini 4 Carbon in Google's Antigravity platform to remake the 1987 roguelike NetHack as a modern 3D action RPG, porting it to web technology and wasm. The build relied on the /swarm feature, which spawns subagents with separate contexts to review each other's output, plus a Blender MCP server for 3D asset generation. Gemini 4 Carbon is available through the Google AI Ultra plan.
LangChain's internal GTM Agent generated 1,540 sales artifacts over 12 weeks, saving reps 337 hours on documents like account plans and proposal decks. 47% of deals with a sent artifact advanced to a later stage, versus 16% of matched deals without one, and 46% of artifacts were followed by a customer call within 14 days. The agent runs on the Deep Agents harness, whose sandbox, skills, memory, and scheduled runs are available in Managed Deep Agents.
AWS shows Java teams how to build autonomous AI agents using Spring AI and Amazon Bedrock's AgentCore, with a spring-ai-agentcore integration. Two reference systems are detailed: a single-agent travel assistant with tool calling and memory, and a multi-agent intelligent document processing platform with five specialist agents, a rules engine, and alerting. The stack requires JDK 21, Spring Boot 4.1 with Spring AI 2.0, and deploys via AgentCore runtime on ARM64 containers.
OpenRouter published a break-even analysis showing when self-hosting an open-weight LLM costs less than calling it through an API, using per-provider prices as of October 8, 2026. At those prices, Qwen3.8 27B on one rented RTX 3090 at $0.50 an hour breaks even at 6 to 12% load, while DeepSeek V4.1 Flash on four H200s never beats the cheapest zero-data-retention endpoint. The post also notes that provider.zdr and the EU and US domains cover most privacy needs without self-hosting, and recommends running a steady base load locally while sending peaks to the API.
OpenRouter published a guide on stopping AI agents that get stuck retrying a failing tool, using four guards tested on six current models. With doomLoop detection enabled, all six models were blocked on their third identical tool call, and across 18 runs tool calls fell from 314 to 64 while spend dropped from $0.29 to $0.05. The guide adds a loopKey to normalize arguments, stopWhen caps on steps and cost, and a credit limit on the API key that returns a 402 with limit_source openrouter_key_limit.
OpenRouter published a guide for benchmarking LLMs on your own application's inputs, grading outputs with code and comparing accuracy, billed cost and request time. On its 60 synthetic tasks, GPT-6.1 Sol passed 180/180 trials across three repeats, while DeepSeek had the lowest response cost at 157/180 passes. Agreement with the public ranking was weak, with a Spearman correlation of 0.109.
vLLM reports serving DeepSeek-V4.1-Flash on a single NVIDIA Vera Rubin NVL72 rack at 1.7M tokens/s per GPU with 387 tokens/s per user, 58x the GB300 NVL72, peaking at 2.52M tokens/s and 1,810 tokens/s for a single user. The gains come from kernels rebuilt for Rubin's split memory domains and FP4 tensor cores, a specialized TP8 decode path, and rack-wide KV and Engram pools shared over NVLink. With prefixes pooled across the rack and Dynamo workload-aware routing, the prefix hit rate holds at 96.5-97% even at high concurrency.
Why it matters: vLLM's third DeepSeek-V4.1-Flash post details how Rubin-specific kernels and rack-wide memory pooling lift throughput 58x over GB300 NVL72, offering a concrete reference for large-scale serving optimization.
Hugging Face published fine-tuning scripts and containers for TRELLIS.2, a 3D generation model, runnable locally or on Hugging Face Jobs. The guide uses LoRA adapters via the PEFT library on only the sparse-structure flow DiT and shape flow DiT, leaving the material stage frozen. On a 188-object footwear subset of the GSO dataset, the pretrained model often rendered incorrect soles, which the LoRA fine-tune targets.