Latest insights

Self-improving training loops generate their own data — and drift when the model also judges it. The versions that actually work add an external verifier, not a bigger model.

Choosing a smaller model trades capability for cost blindly. Compression keeps the model that passes your eval and trims what it doesn't use — but only an eval tells you how far you can cut.

In a multi-agent system, a locally perfect prompt can make the whole worse. The fix is to optimize prompts for the handoff, not the agent — the same reason these systems are hard to debug.

Prompt injection can't be filtered away — instruction and data share one channel. The defense isn't a smarter prompt; it's least privilege on the tools the agent can call.

Most agent evals score the final answer and ignore the tool calls. But an agent can be right by luck or by an unauthorized action — score the tool use separately, or it bites you in production.

What an AI agent actually costs to build — real ranges, the hidden total-cost-of-ownership nobody quotes, and why we publish our prices when almost no one else will.

AI coding agents are only as good as the context they load. A small, current, in-repo wiki written for the machine makes an agent fit your codebase instead of fighting it.

A new paper shrinks model hallucination again — but no technique makes it zero. The production answer isn't a fix; it's the eval, the gate, the trace, and the checkpoint.

When a boutique AI firm beats the Big 4 and when it doesn't — the honest split on cost, seniority, and who actually builds the system you're paying for.

60–95% of AI pilots never reach production. The cause is almost never the model. Here are the real reasons they die, and the three operating rules that keep ours alive.

A practical checklist for deciding whether an AI agent is safe for production — building or buying. The seven things that separate a demo from a system.

The warning signs that an AI vendor will burn your budget — and the questions that separate a firm that ships production systems from one that ships a demo and a roadmap.

Multi-agent systems fail in ways event logs can't see. The fix is a trace — a typed tree of spans with the reasoning attached. Observability is part of the build.

When agents adapt from production feedback, one can improve locally while breaking what its peers depend on. Version-stamped traces catch the drift before users do.

Proprietary tracing formats are a lock-in tax. OpenTelemetry's GenAI conventions let you instrument an agent once and send the traces anywhere you choose.

A field guide to agent observability on Azure — Semantic Kernel, Azure AI Foundry tracing, and Application Insights, kept portable with OpenTelemetry and no lock-in.

A field guide to agent observability on Google Cloud — ADK and Agent Engine on Vertex AI, Gemini, and Cloud Trace, kept portable with OpenTelemetry.

The difference between a chatbot and an agent system is the difference between a tool and a team. What an agent system actually is, and why it is the work that lasts.

Faceless AI content channels are a business model with no quality bar, and they fail for the same reason ungated AI systems fail in production. The eval is the difference.

AI content repurposing is real demand with a closing window — the tools are automating the service away. A way to think about building where the moat outlasts the model.

Anyone can deploy a website chat agent in an afternoon. That is exactly the problem. Where these quietly fail, and why the answer is a system, not a widget.

The voice-agent demo is an afternoon. The system that answers a plumber's phone at 2 a.m. without losing the job is the actual work. Here is the gap.

A field guide to AI snake oil — the trading bots, the guaranteed returns, the products that are really courses. The test is simple, and it generalizes.

Most AI consulting fails before the build, in a vague assessment nobody can act on. The audit is not a sales step. It is the first deliverable, and it should stand on its own.

AI-assisted coding is becoming table stakes. AI systems engineering is becoming the real differentiator. Here is the difference, and why it matters.

Tool-using agents need an explicit map of what they can read, write, mutate, escalate, and never touch — decided before the build, not during an incident.

Before you tune retrieval, prove the question can even be answered from the source corpus — with a citation a human would accept. Answerability comes before embeddings.

Production AI fails in ways ordinary runbooks don't cover. The operating plan must handle quality drift, retrieval failure, model outages, cost spikes, and human escalation.

Prompt caching is no longer a performance optimization. It is an architectural constraint that decides whether a long-running agent is economic to operate.

The Model Context Protocol is moving from a developer convenience to the production interface between agents and your systems. Here is what changes when you treat it that way.

A small verifier model sitting between the frontier model and the side-effect boundary is the most useful piece of agent architecture nobody is shipping yet.

How a working eval harness picks the model — and how often the cheapest model that passes beats the frontier one the team came in expecting.

A PoC with no path to production hides the hard decisions. A week-2 user forces them into the open while the architecture is still cheap to change.

Why a real user belongs inside the system by week 2 — and the kinds of architecture decisions they force into the open that no spec would catch.

How a discovery eval decides whether a build should happen at all — and why the honest output is sometimes 'don't build', before you've spent the build budget.

An evaluation harness is a product, not a notebook. Ship it like one — versioned, owned, instrumented, and able to stop a release when a run goes red.