Tag: ai-agents
Do you need a director of ontology for AI agents?
A director of ontology builds the shared vocabulary AI agents reason over. Capital One, Novartis and Lockheed Martin are hiring for the work, yet UK ads citing ontologies were 0.035% of permanent job ads in the six months to October 2026. Most companies should start with typed files instead.
The LLM writes, the model decides
Pair a writing LLM like Claude with a cheap decision model that only returns numbers, such as TypeSafe Jev. Three ways to split the work, why a probability Claude writes down is a self-report and not a measurement, and why a reason written after the fact for a bare number is worse than no reason.
How to run many Claude Code sessions without duplicate work
Claude Code sessions can now message each other over a Unix socket on your own machine. I had 23 sessions running and only 14 of them could see each other. Here is what the message channel fixes, what it cannot fix, and why the guard I trusted for months was never a guard at all.
Two of your agents never read your CLAUDE.md
Claude Code loads your CLAUDE.md into every subagent except two. Explore and Plan skip it by design, and no setting changes that. So the agents you fan out most widely are the ones that never see your rules. ETH Zurich measured what the file costs on the occasions it does load.
How to run a long autonomous Claude Code job without it drifting
The hard part of a big AI job is not the work. It is making the agent run for many sessions without drifting or claiming it is done when it is not. I used an accessibility audit across four codebases as the test. The setup that kept Claude Code on track was a git ledger, atomic parallel claims, and two verification passes.
A report and a semantic model are not the same thing, and your AI agent only cares about one of them
Most people treat a Power BI report and its semantic model as one object. They are two files doing two jobs. When you point an AI agent at your data, the report is the cheap half and the semantic model is the part that took three years to get right.
BI only ever saw half your company. AI can see the other half
Business intelligence was always the quantitative side: rows, numbers, things that fit in a column. The qualitative half, the calls and emails and tickets where the why actually lives, was invisible to it. That half is most of your data, and it is where AI adds value BI never could.
Managed AI agents and the cost crossover nobody calculates
Anthropic managed agents bill $0.08 per session-hour, and everyone races to compare that to a cheap VM. The comparison misses the point. Runtime is a rounding error next to tokens, and the operations bill decides the rest. Here is where self-hosting an AI agent actually starts to pay, with the real 2026 numbers.
One-time question or a permanent dashboard? AI just changed the answer
Every BI team has quietly run the same triage for years: is this worth a dashboard, or is it a one-off? Building a dashboard was the only durable option, so the long tail of one-time questions mostly went unasked. AI collapses the cost of the one-off, and that reshapes the whole portfolio.
Should you build your agents in Copilot Studio? The demo is not the question
Low-code agent builders like Copilot Studio get you to a working demo in an afternoon. That is real, and it is also the trap. The question is not whether it demos well. It is what you give up the day you need control, and whether you will need control.
Stop building "the ERP agent." The value is in skills that cross functions
Companies build AI agents shaped like their org chart: an ERP agent, an HR agent, a finance agent. Each one is a silo with a chat box. The real payoff shows up when skills compose across functions, because data exists to tell a story or trigger an action, not to sit in one department.
The hard part of analytics AI is not the answer, it is figuring out the question
Everyone obsesses over whether the model reasons well. The real failure in AI over business data happens earlier, at the moment the agent decides what you meant. A confident answer to the wrong question is worse than no answer at all.
Your old dashboards are the answer key for your new AI
Teams building analytics AI keep starting from a blank page. Meanwhile the most validated business logic they own is sitting in the dashboards they already shipped. Those reports are years of distilled definitions and a ready-made test set. Mine them.
When to use a dynamic workflow
A dynamic workflow in Claude Code runs up to sixteen subagents at once and a thousand across a job. That power is wasted on most tasks. This is the decision I use before reaching for one: when a single agent wins, when a dynamic workflow earns its cost, and when the answer is to not automate at all.
AI does tasks. It does not do jobs.
Ten years building Tallyfy, and a year pointing AI agents at it, taught me one blunt thing. A job is a chain of tasks, and AI reliability multiplies down that chain until the whole thing is a coin flip. The fix is not a smarter model.
Dynamic workflows: parallel verification at scale
Dynamic workflows in Claude Code run tens to hundreds of subagents that check each other before anything reaches you. The parallelism is not the interesting part. The verification is. Here is how I am using one to re-verify around 300 posts on this site, and when it earns its cost.
Anthropic managed agents are not office agents
Anthropic managed agents and office agents are different products with confusingly similar names. Managed agents is a developer API for running autonomous Claude agents on managed infrastructure. The interesting part is the brain-hands split: Anthropic runs the agent loop, while the sandbox can run in your own environment. This is what it is, and when to use it.
How the general-purpose agent works in Claude Code
The general-purpose agent in Claude Code is not the main agent and not something you pick. It is a built-in subagent Claude routes to on its own for complex, multi-step work. It inherits your model and, by default, runs in its own fresh context that Claude briefs with a short summary. This post explains how it actually works and what that costs you.
How to debug Claude Code subagents
When a Claude Code subagent fails, you cannot open it and look inside. It ran in its own isolated context and handed back a summary. Debugging a subagent is the skill of reading that summary, recognizing context-isolation failures, and designing subagents that report enough to be diagnosed. Here is how to do it.
Self-hosted vs managed AI agents is a governance call
The choice between self-hosted and managed AI agents gets treated as build versus buy, a cost question. It is not. It is a governance decision about where your data goes, what you can audit, and whether you can leave. Here is what each path gives you and how to decide.
The built-in agent types in Claude Code
Claude Code ships with five built-in agent types: Explore, Plan, general-purpose, statusline-setup, and claude-code-guide. Most people know two of them. The other three run constantly and shape how much your sessions cost. This is the full catalog, what each one is for, and why knowing them changes how you read your own terminal.
What is a subagent in Claude Code
A subagent in Claude Code is a specialized worker that runs in its own fresh, isolated context window, with its own tools and permissions, and reports back only a summary. It is how Claude does a noisy side task without flooding your main conversation. Here is what a subagent is, what file defines it, and when it earns its cost.
Subagent vs parallel agent vs skill in Claude Code
Subagent, parallel agent and skill get used as if they mean the same thing in Claude Code. They do not. A skill is reusable instructions that cost almost nothing until invoked. A subagent is delegated work in a fresh isolated context. Parallel agent is not a primitive at all. Picking the wrong one wastes tokens or floods your context.
What Claude office agents actually do and why you should care
Claude office agents let Claude share context across Excel and PowerPoint through a single toggle. Here is what the feature actually does, what the Skills framework changes, and the security gaps you need to know about before enabling it.
Multi-agent orchestration - the complexity trap
Multi-agent AI systems promise specialized intelligence but deliver exponential complexity. Salesforce research shows agents achieve only 58 percent success on single tasks and adding orchestration doubles the failure rate. Most mid-size companies need one capable agent, not coordinated swarms.
Using Claude Code for legacy modernization - 90 days does not finish it, but proves it is possible
Stop thinking 90 days will complete your COBOL to cloud migration. Utah took 18 months and AWS Transform cut Toyota timeline by 50%. Use that time to prove legacy modernization works, build organizational confidence, and create momentum for the multi-year migration ahead.
Agentic AI use cases that actually work
A large share of agentic AI projects face cancellation due to poor problem selection. Companies like Ciena and IBM show where autonomous agents deliver real value. Here are the specific use cases that work and when to skip agents.
Building reliable AI agents - why boring beats brilliant
OpenAI GPT-4o failed 91.4 percent of office tasks in testing. Reliable AI agents require engineering discipline over model brilliance, with proven patterns like circuit breakers and error budgets that turn prototypes into trusted production systems.
Building your AI roadmap: the template
Most AI roadmaps focus on capabilities and features when they should focus on reliability and failure modes. RAND Corporation found more than 80% of AI projects fail before production, and only a small fraction of organizations have scaled AI fully across the enterprise. Your roadmap must prioritize reliable agent patterns over impressive demos. Start with constraints, measure operational health, and plan for continuous iteration.
OpenAI Assistants API: the good, bad, and expensive
OpenAI Assistants API packs stateful conversations, code execution, and document search into one package. Built production systems with it and found the complexity rarely justifies the cost. With deprecation complete as of August 2026, here is when it was worth using and when simpler alternatives win for chatbots and automation.