Hightower's AI Harness Engineering
Articles about Hightower's AI Harness Engineering
Price Versus Performance: The Frontier SI Models of October 2026, Haiku 5.5 is a barn burner
Every frontier lab now sells several models at different price points, and the sticker price alone says little about what you get.
AI News, Volume 50: The Small Model Moves Into the Machine Room
Claude Haiku 5.5 cuts the cost of agent swarms as Microsoft, Google, Perplexity, OpenAI, GitHub, and Cloudflare standardize skills, hooks, sandboxes, and recovery.
AI Agent Eval Market 2026
Evaluation Stacks That Survive the Market: Cleaned Transcript
Eval Agent Market Consolidation
ClickHouse acquired Langfuse. Cisco acquired Galileo. OpenAI absorbed Promptfoo. And Dynatrace signed a deal for Arize.
AI News, Volume 49: Always-On Agents Get a Kill Switch
OpenAI launches dots, computer-use APIs, GPT-6.1 Sol, and event-driven plugins; NVIDIA moves safety below the model while Nasdaq brings governed MCP into capital markets.
Microsoft Foundry HyperScaler VII: Microsoft Foundry Ships a Thousand Tools. The Only Two Questions That Matter Are Not on the List
Part 7: A managed tool catalog is not interesting for the tools it has. It is interesting for the decision rule behind each one, the constraint that bites in production, and the governance layer
AI Eval Market 2026
Evaluation Stacks That Survive the Market
AI Agent Hyperscalers: Microsoft Foundry vs Amazon Bedrock AgentCore
Two hyperscalers now rent the same agent substrate. This paper compares them layer by layer and says where each one actually wins.
AI News, Volume 48: The Model War Moves Into Operations
Anthropic ships Sonnet 5.5, OpenAI launches Sol and Luna, xAI pairs Grok 4.7 with Team Bots, and two security failures redraw the agent operating boundary.
LangFuse VI: Stop Shipping Prompt Wording in a Pull Request
Part 6: Hardcoded prompts turn a two-minute wording tweak into a pull request. Langfuse prompt management fetches the production label at runtime so wording ships without engineering.
Microsoft Foundry HyperScaler VI: Your agent’s tools should not ship on your release cadence
Part 6: The real argument for a managed tool registry is not convenience. It is that authentication stops being code, and tool policy stops sharing a release cadence with your agent.
Validator-First Neuro-Symbolic Architectures for Deterministic Production Agentic AI
Validator-First Neuro-Symbolic Architectures for Deterministic Production Agentic AIA production agent cannot treat a sampled model output as the last word on whether an action runs. The model proposes. A typed, re-runnable check decides. The split is the architecture. The scores and product names below are supporting evidence, not the product.
Practitioner Paper: Implementing the Validator-First Loop: Policy Engines, Shapes, Solvers, and Credentials for Neuro-Symbolic AI
Part 2: Refund loop can move to a specialized, versioned artifact: SHACL for shape, Rego or Cedar for policy, PAL for arithmetic, a Z3 theory for a solver-backed claim, and Open Agent Passport, etc.
Validator-First Neuro-Symbolic Architectures for Production Agents
A production agent cannot treat a sampled model output as the last word on whether an action runs. The model proposes. A typed, re-runnable check decides. The split is the architecture.
Microsoft Foundry HyperScaler V: Checkpoint your current state and save it in the Azure AI Hyperscaler
Part 5: Foundry ships four state services with four different lifetimes, and the checkpoint class everyone recommends for wiring them together doesn’t exist; here is the right mental model
Microsoft Foundry HyperScaler IV: Your Foundry agent is deployed and has no idea who is calling it
Part 4: The Foundry-hosted agent runtime contract looks like a table of transport trivia, but the two request headers at the end decide whether your multi-tenancy is real.
Microsoft Foundry HyperScaler III: Every platform says it is Framework-Agnostic. Here is the receipt
Part 3: Every vendor claims to be framework-agnostic, so run the controlled experiment: rebuild the same agent in LangGraph, deploy it with the identical five commands, and audit the four places
Microsoft Foundry HyperScaler II: Your agent is four lines away from a governed cloud endpoint
Part 2: Microsoft Azure’s HyperScaler - Foundry Agent Service. The lift from a laptop agent to a governed, identity-bearing Foundry endpoint is four lines of hosting code, one YAML file, and
Microsoft Foundry HyperScaler I: Is Not a Product You Log Into. It Is an ARM Resource
Part 1: Microsoft Azure’s HyperScaler - Foundry Agent Service. You already know how to write an agent loop. The question nobody answers honestly is what the platform actually owns.
The Azure Microsoft Hyperscaler Control Plane
A series (plus four appendices) on keeping your agentic loop and renting the substrate: Microsoft Foundry, Microsoft Agent Framework, and LangChain DeepAgents.
The Model That Cannot Write a Sentence: Jev System One Saves 26x to 260x and can be 20x to 200x faster
A yes-or-no question does not need a language model. It needs a calibrated probability. Perfect for adversarial subagents, judges, and agentic checkers. Oh, and it is way faster too!
AI News, Volume 47: Agents Get Memory and a Watchtower
Google monitors anomalous agents and opens the smart home to MCP, Anthropic folds Cowork into Claude, Perplexity moves onto HP workstations, and xAI makes memory a default.
The Second Brain Compared with Other Agent Memory Systems
Using OKF, custom ontologies, to define a second brain to capture SLDC, system architecture, decisions, compared to other agentic memory systems. Git-native institutional memory vs. Mem0, Letta, etc.
The Hosted Runtime: Loop Engineering and the Harness Around It IV
What survives when OpenAI runs the Codex harness for you
The Judge Said the Same Thing Four Times: Harness Engineering Part XI
Part 11: A judge that wobbles announces its own unreliability. A judge that repeats itself does not; my LLM-as-judge returned four byte-identical verdicts on a figure that was fine.
Agent Memory
Agent Memory as Harness Engineering: Retention, Eviction, and Reload in Production Agent Loops
AI News, Volume 46: Agents Leave the Lab
A $12.93 billion acquisition meets the managed-agent era: NVIDIA buys Hugging Face, OpenAI launches Astra and the Agents API, Meta ships Muse, and Google locks down the sandbox.