AI costs · Evals · Production engineering
Independent FDE perspectives, model updates, and practical guides to costs, evaluations, and production AI.
All articles
Newest first
A researched editorial shortlist of ten AI transformation providers, with Deployed Engineer featured first as the publisher's practice, service comparisons, and questions to ask before choosing a partner.
A 2025-informed editorial shortlist, published in 2026, featuring Deployed Engineer and nine providers, with six clearly hypothetical industry case studies.
A researched guide to TypeSafe's Jev with six classification workflows, a Python ticket-routing example, and a practical evaluation plan.
The opportunity in AI services sits between model capability and everyday operations. Here is how an independent forward deployed engineer closes that gap.
GPT-6.1 Sol, Grok 4.7, Gemini 3.8 Flash, and Claude Opus 5.5 bring new options for production workflows. A source-linked briefing on what changed and what to evaluate.
A first-principles playbook for cutting LLM and AI infrastructure costs by measuring cost per successful task, routing work, caching context, and using evals as the safety rail.
A beginner-friendly guide to testing a support chatbot with five questions, simple pass-or-fail checks, and a repeatable spreadsheet workflow.
A simple guide to checking whether a chatbot chose the right tool, used the right details, asked for confirmation, and handled failures safely.
Test whether a RAG chatbot found the right passage, answered from that passage, cited it, and admitted when the answer was missing.
A practical, source-linked guide to GPT-6 Astra's long-context reasoning, computer use, coding, and professional workflows—with a cost-aware way to decide when it is actually the right model.
A step-by-step demo for building a support-triage agent with tools, a sandbox, approvals, subagents, and traces—without hiding the important engineering decisions behind a framework.
A source-checked briefing on the launches moving agents from chat windows into browsers, developer infrastructure, and long-running company workflows.
When agents can generate more code than a person can read, understanding has to come from traces, executable examples, diagrams, and sequential explanations.
A practical reading of Gamma's reported AI visibility tactics, including first-party content, agent-readable actions, comparison pages, and citation monitoring.
Cursor is expanding from editor to forge while Command Code learns from developer choices. Here is what changed, including the verified Grok model and current pricing.
A source-checked model briefing on Qwen3.7 Flash, Qwen-Image 3.0, GLM-5.3, and the model name that Google has not announced.
Studio by Spotify Labs can research, organize work, use connected context, and turn it into personal audio. That makes taste part of the agent interface.
Unsloth brings local model use, comparison, tool calling, file analysis, and training into an offline desktop workflow for Mac and Windows.
Anthropic explains how session habits affect token use while OpenAI cuts GPT-5.6 Luna and Terra prices. Both point toward cost per successful task.
A YC launch for Pentagon puts a useful name on an emerging practice: agent teams need explicit, testable organization design.
My organized notes from AI Engineer World's Fair sessions on July 1 and July 2, covering computer use agents, eval design, agent simulations, CaaS, AI-first SDLC, LLM wikis, and Ramp's GTM agent platform.
Notes from Leo Mehr on Ramp's FDE practice: always be scoping, interrogate urgency, validate assumptions, and use AI to speed up request triage.
Notes from Kevin Bai on Anthropic's FDE framing: sell outcomes, build on shared primitives, and only add FDE when a complex product meets a non-technical buyer.
Anthropic launched Claude Fable 5 on June 9, suspended access on June 12 after US export controls, then restored global access on July 1. Here is what changed, what the safeguards mean, and where Fable fits in serious agent stacks.
Z.ai released GLM-5.2 with a solid 1M-token context window, MIT-licensed weights, 744B total parameters with 40B active, and strong long-horizon coding results. The open-weight agent stack is getting serious.
Vercel launched eve, an open-source framework for production agents. Instructions live in Markdown, tools in TypeScript, skills and subagents in folders, with durable execution, sandboxed compute, approvals, and evals built in.
Loop engineering moved from X posts into serious engineering blogs in June. LangChain's stacked-loop framing makes the idea more concrete: agent loops, verification loops, event-driven loops, and trace-driven improvement loops.
Anthropic shipped Claude Opus 4.7 this week. 87.6% on SWE-bench Verified, 64.3% on SWE-bench Pro, a 1M-token context window, and the same $5/$25 pricing as Opus 4.6. Here is my honest take on what changed, what did not, and whether it deserves a spot in your stack.
OpenAI is wiring a Scratchpad UI into the Codex app that fires multiple agents in parallel from a TODO list. A small feature with a big signal. Parallel agent execution is no longer a differentiator. It is the default.
Google released Gemma 4 on April 2 under Apache 2.0. Four sizes from 2B to 31B, 256K context, native vision and audio, and real agentic capabilities. The open-model landscape just changed, and not everyone is ready for what that means.
OpenClaw is an open-source autonomous AI agent with 310K+ GitHub stars. Learn how it bridges messaging apps and AI models to automate tasks locally — and why agent engineers should care.
OpenAI shipped GPT-4.1, o3, o3-pro, and o4-mini. Here is what each model does, when to use it, and how they change the AI engineering stack.
The 2026 guide to Claude Code: voice mode, /loop, skills system, 1M context, Opus 4.6, and the workflows that top AI engineers actually use.
Cursor 2026 adds event-driven automations, parallel cloud agents, Bugbot Autofix, and MCP Apps. Here is what changed and why it matters for AI engineers.
A complete guide to MCP directories with 1200+ servers for AI engineers building with Claude Code and Cursor. The Model Context Protocol ecosystem mapped.
A direct comparison of the 5 best AI CLI tools for developers in 2026 — pricing, capabilities, and which to choose for your engineering workflow.
A forward deployed engineer (FDE) embeds directly with customers to solve their hardest technical problems. Learn about the role's origin at Palantir, why it grew 800% in 2025, salary data ($238K-$630K+), required skills, and how to break into this career as a forward deployed AI engineer.
Master Cursor AI with these viral tips that are transforming how engineers write code. From prompt engineering to code generation, learn the secrets that top developers are using.
Discover the viral Claude coding techniques that are revolutionizing how developers approach problem-solving and code generation.
Curated list of the best learning resources for engineers in 2025. From AI tools to system design, discover what top engineers are using to stay ahead.