Troy Magennis
Bellevue, Washington, United States
5K followers
500+ connections
About
Troy Magennis is a seasoned technology leader, consultant, and entrepreneur with over 25…
Services
Articles by Troy
-
Will AI Change Developer Metrics?
Will AI Change Developer Metrics?
[ This was my most recent newsletter article. If you want to subscribe, use the form on my home page…
7 Comments -
LLM as a Service Product Updates (28-Apr-25)
LLM as a Service Product Updates (28-Apr-25)
YOUR ACTION ITEMS 1. Model updates: It's been a busy month of new vendor Model releases, led mainly by OpenAI.
-
Predictability and Flow Accelerator Consulting Package (5 days)
Predictability and Flow Accelerator Consulting Package (5 days)
I've been consulting and training on predictability and forecasting for over 20 years. Only recently did I feel…
5 Comments
Activity
5K followers
-
Troy Magennis shared thisYou can get a lot of value without exact measurements. Simple scoring and reflection often lead to learning. For example, here is a simple scoring for post-release retrospection on features chosen for delivery. The goal is to keep the learning in the AI decision-making loop.
-
Troy Magennis posted thisHow do we measure AI performance? I've been giving this some thought, but I needed a set of principles to guide the right measures. Here is what I have so far, but what's missing? 1. Output is not outcome 2. Speed without direction is waste 3. Work should be judged comparatively, not in isolation 4. Cost of delay matters 5. Risk gates value 6. Delivery cost is broader than coding cost 7. Review is a learning system 8. AI productivity must be measured at the system level 9. Strategy must be concrete enough to order choices 10. A healthy system learns Deep-dive on each here: https://lnkd.in/gs7q92Cm I stopped there. And made sure the metric dimensions I came up with aligned to one or more of these. If you want to deep-dive on each, take a look as Larry Maccherone and I are building this in public, and it's VERY VERY much in a state of flux (we are still even trying to determine if we have the right starting point!) - but please join us! It will be fun!
-
Troy Magennis shared thisI know many of you have "real lives," but for those like me who don't, about 15 years ago, I looked at different voting systems and implemented many algorithms. There are a heap of ways voting can be intentionally and unintentionally gamed (going against the group consensus). If you want to experiment or need to do a "fair" ballot, here is a page for you. https://lnkd.in/gXghyRyu The real art form is handling people's preferences beyond their favorites. with rankings of "preferred" and "equally likable to me" - If in doubt, use the Schulze Method. Vote on my friends....
-
Troy Magennis shared thisDo you have a calendar booking tool? I'm got my AI "friends" to build one of my own for my website. Everyone seems surprised that I'll hop on a call with them to chat metrics, forecasting, and dependencies. So help me out, book some time to test out the page: https://lnkd.in/gPJmnCiE And I'll spend some time talking metrics stuff with you. Deal?
-
Troy Magennis shared thisNew video series: AI App LLM Cost Savings — practical engineering patterns for keeping AI features fast, reliable, and economically sustainable in production. Sixth topic: Batch Mode Is the Overnight Discount https://lnkd.in/gkF7CtJ8 Batch mode is one of the easiest ways to reduce AI application costs when the user does not need an immediate response. In this video, I explain how batch APIs can act like an “overnight discount” for LLM work by trading real-time response guarantees for lower input and output token costs. I cover which use cases are good candidates for batch processing, such as nightly reports, CRM enrichment, document tagging, summarization, analysis, and backfills, and which tasks still need real-time responses, such as fraud checks, live assistants, or user-facing decisions. The key pattern is simple: use real-time AI only when someone is waiting, and move everything else into asynchronous or batch workflows where possible. I’d love feedback from developers building AI features: where are you seeing cost creep in your apps?6 - Batch Mode is Your Overnight Discount (Reducing AI Cost Series)6 - Batch Mode is Your Overnight Discount (Reducing AI Cost Series)
-
Troy Magennis shared thisNew video series: AI App LLM Cost Savings — practical engineering patterns for keeping AI features fast, reliable, and economically sustainable in production. Fifth topic: Stop Paying Your Model to Think https://lnkd.in/gtQn5tDu Reasoning models are powerful, but not every AI call needs the model to “think” deeply before answering. In this video, I look at how reasoning levels can become a hidden cost driver in AI applications, especially when simple tasks like classification, extraction, formatting, or short summarization are sent through high-reasoning models. I cover why developers should choose the lowest reasoning level that still produces reliable results, turn reasoning down or off in development environments, test whether higher reasoning actually improves quality, and reserve deeper reasoning for complex, high-risk prompts where the extra latency and cost are justified. I’d love feedback from developers building AI features: where are you seeing cost creep in your apps?5 - Stop Paying Your Models to Think (Reducing AI Cost Series)5 - Stop Paying Your Models to Think (Reducing AI Cost Series)
-
Troy Magennis shared thisNew video series: AI App LLM Cost Savings — practical engineering patterns for keeping AI features fast, reliable, and economically sustainable in production. Fourth topic: Prompt Caching Only Works if Your Prompt Stops Changing https://lnkd.in/g5-3tZM6 Prompt caching can reduce AI application costs, but only if your prompts are structured consistently. In this video, I explain why stable content should come first, variable content should come last, and why small changes like timestamps, reordered tools, duplicated boilerplate, or moving JSON schemas below runtime data can reduce caching effectiveness. I also show how shared prompt templates, consistent system instructions, reusable tone and policy files, and predictable prompt composition can help developers get more benefit from provider-side prompt caching without relying on every team member to remember the rules manually. I’d love feedback from developers building AI features: where are you seeing cost creep in your apps?4 - Prompt Caching Only Works if Your Prompt Stops Changing (Reducing AI Cost Series)4 - Prompt Caching Only Works if Your Prompt Stops Changing (Reducing AI Cost Series)
-
Troy Magennis shared thisNew video series: AI App LLM Cost Savings — practical engineering patterns for keeping AI features fast, reliable, and economically sustainable in production. Third topic: Your Context Window is Your Biggest AI Bill Your context window may be one of the highest hidden costs in your AI application. As teams add longer system prompts, chat history, retrieved documents, tool definitions, safety policies, JSON schemas, and runtime application data, every model call gets larger, slower, and more expensive. In this video, I look at why “the cheapest token is the one you never send,” how oversized context can hurt both cost and response quality, and practical ways developers can control it: setting context budgets, trimming unnecessary inputs, summarizing long histories, limiting retrieved documents, and only exposing the tools a prompt actually needs. https://lnkd.in/gmCqBb8a I’d love feedback from developers building AI features: where are you seeing cost creep in your apps?3 - Manage Your Context Window (Reducing you AI Cost Series)3 - Manage Your Context Window (Reducing you AI Cost Series)
-
Troy Magennis shared thisI just cancelled GitHub Copilot. I've been with them from the start. They are moving to token usage based pricing, which I understand, but their pricing model makes a huge mistake. it's a pay upfront, and use it or lose it at the end of each month. If I'm going to pay API prices, why doesn't it roll over? it's a miss. I've bet big usage based pricing is going to be essential in the future. and even modeled different strategies at https://usagetap.com so I've got strong views. here are the mistakes I see: 1. Either allow people to purchase credit in advance, or do expiring subscriptions. but don't penalize people for paying in advance. 2. If you allow long running agentic tasks, you need safeguards against abuse. but don't punish the normal human users (me). I never reached my limit. 3. Model your pricing strategy for different tiers and types of users. One size doesn't fit all. Am I being irrational unfair? please comment below. I'm doubling down on my CODEX subscription. Which I'm finding the cloud option great, as I can code while hiking!!!
-
Troy Magennis liked thisTroy Magennis liked thisI realised during Troy’s presentation I was capturing every slide. So it’s impossible to post that here. But every slide was sparking more and more thought how will AI disturb our world. It’s hot right now in all news that coordinating and aligning on stuff will become the hardest task. If we can now start anything we want, why shouldn’t we? So imagine the organisations who already have a very poor visualisation of what they are working on, starting more will lead to disaster. The question what should we start next and when will have more weight than ever. Powerful thinking, dependencies and metrics. Love it. #FLDay #alignement #dependencies #metrics
-
Troy Magennis liked thisTroy Magennis liked this🤖✈️ AI & Flight Levels Excerpts from Flight Levels Day 2026 I attended Flight Levels Day last night and two sessions stood out on the AI topic. Troy Magennis analyzed how AI changes dependencies in Flight Levels systems. His finding: AI dramatically reduces Knowledge and Task Dependencies — but makes Alignment Dependencies worse. When coding takes hours instead of weeks, the bottleneck shifts to everything else: legal sign-off, marketing, training, cross-platform release coordination. FL2 must now “double down” on coordinating the non-coding work. And FL1’s new job? Protect downstream resources to maximize flow — not just start whatever has the highest impact. Thomas Krause interviewed six pioneers across industries, asking about AI. They answered with Org Design questions. A recruiter who built 100K lines of code in 8 weeks stopped and said: “A tool alone does nothing. We need end-to-end flow.” An agile coach in a 125-year-old energy company isn’t deploying AI yet — she’s building coordination-level interactions first. Thomas argues this groundwork is the prerequisite: organizations that understand their value flow will be ready when AI knocks on the door. Those that don’t will just build the wrong thing faster. Thomas also proposed three moves: 1️⃣ Descend & Discover Value Flow before inserting AI agents 2️⃣ Rethink Processes instead of just automating existing ones 3️⃣ Treat Work Systems Design as a continuous practice (Adaptive Operating Model) — not a one-time transformation. Both speakers converge on the same conclusion — and it echoes what Klaus Leopold has been saying: most AI use cases today help individuals (FL0). The real leverage is at FL2 and FL3. AI accelerates execution, but coordination and strategy don’t speed up by themselves. Without Flight Levels (or something similar to this nature), you just get the Disappointment Loop of building the wrong things — faster.
-
Troy Magennis liked thisThe vast majority of my work involves helping companies unlearn and fix mistakes they have done in the past. This is often slow, expensive, painful and riddled with wicked challenges. So, here is a radical thought: What would happen if we helped companies avoid those mistakes in the first instance? For example, a key moment in a company's history is when it is growing past its early startup stage. It's at that point in time when companies often start adding hierarchical structures to handle their growth. Later on, they end wondering how their startup culture and dynamism vanished! The next Business Agility Finland event with Juhani Tapaninen, Eliška Čejpová and I is tomorrow (Thursday 23 April) and we will be exploring how companies can avoid common pitfalls of growth. It's organised in partnership with the awesome people at Helsinki Incubators.Troy Magennis liked this📈Is your startup ready for its next growth phase? If the answer is yes, then you should join us on April 23 for "Avoiding the Pitfalls of Startup Growth," a workshop hosted by Business Agility Finland. This session is designed for founders and leaders looking to grow their teams effectively without losing the dynamic energy that made them great in the first place! During the workshop, we’ll help you spot common "growth traps" and successfully navigate the shift to a mid-sized company. Here is the lovely team hosting the event: Juhani Tapaninen, Eliška Čejpová, and Jose Casal 🤝💙 They’ll be using "Training from the BACK of the Room!" principles to ensure you leave with a strategic plan, not just theory. 👉Details: When: 23 April, 17:00-19:30 Where: Helsinki Incubators Space, Porthania 3rd Floor Find the sign-up link in the comments, and we hope to see you soon!🌟
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Publications
-
Modeling, Simulation, and Data Mining
Agile 2012
Modeling and simulating provides and avenue for quickly answering tough management questions about delivery forecasts and performance. Modeling and simulating potential project outcomes hundreds of times (using estimate ranges rather than single values) achieves a higher degree of outcome certainty leading to accurate forecasts, and well-informed management decisions. Finding insight from cycletime and other metrics that can be automatically extracted from lifecycle management tools makes…
Modeling and simulating provides and avenue for quickly answering tough management questions about delivery forecasts and performance. Modeling and simulating potential project outcomes hundreds of times (using estimate ranges rather than single values) achieves a higher degree of outcome certainty leading to accurate forecasts, and well-informed management decisions. Finding insight from cycletime and other metrics that can be automatically extracted from lifecycle management tools makes modeling and simulation more accessible than ever before.
Other authors -
-
Forecasting and Simulating Software Development Projects: Effective Modeling of Kanban & Scrum Projects using Monte-carlo Simulation
See publicationForecasting and Simulating Software Development Projects explains how to effectively model Kanban and Scrum projects to get accurate forecasts of cost, delivery dates and staff requirements.
Modeling using Monte-carlo simulation allows rapid what-if analysis to find options that minimize cost and delivery time, whilst maximizing revenue. Simulation lets you hit target delivery dates, and shows the impact of hiring (or losing) staff with certain skillsets, taking software project…Forecasting and Simulating Software Development Projects explains how to effectively model Kanban and Scrum projects to get accurate forecasts of cost, delivery dates and staff requirements.
Modeling using Monte-carlo simulation allows rapid what-if analysis to find options that minimize cost and delivery time, whilst maximizing revenue. Simulation lets you hit target delivery dates, and shows the impact of hiring (or losing) staff with certain skillsets, taking software project leadership to a new level of maturity. -
LINQ to Objects Using C# 4.0: Using and Extending LINQ to Objects and Parallel LINQ (PLINQ)
Addison-Wesley
See publicationYour Complete Example-Rich Guide to Using and Extending LINQ to Objects and PLINQ.
Drawing on his unsurpassed experience coding software using LINQ and hosting the popular HookedOnLINQ.com site, Magennis presents timely, realistic solutions to a wide spectrum of development challenges, such as combining data from multiple sources, and achieving optimal performance with extremely large sets of data. You’ll begin with brief quick-starts that walk you through LINQ to Objects’ key features…Your Complete Example-Rich Guide to Using and Extending LINQ to Objects and PLINQ.
Drawing on his unsurpassed experience coding software using LINQ and hosting the popular HookedOnLINQ.com site, Magennis presents timely, realistic solutions to a wide spectrum of development challenges, such as combining data from multiple sources, and achieving optimal performance with extremely large sets of data. You’ll begin with brief quick-starts that walk you through LINQ to Objects’ key features and query syntax. Next, you’ll drill down to detailed techniques for applying and extending these features with C# 4.0 and C# 3.0—including code examples that reflect the realities of production development.
Coverage includes
• Writing basic LINQ queries with C#: filtering, projecting, and sorting data from in-memory collections
• Mastering advanced techniques for grouping and joining data and understanding the performance implications of each technique
• Taking full advantage of LINQ’s standard query operators
• Creating custom query operators that follow best practices for coding patterns and error handling
• Writing more fluent, readable LINQ queries with C# 4.0’s language enhancements, including the new dynamic features
• Combining LINQ with COM-Interop to access data sources such as Microsoft Excel
• Using Parallel LINQ to Objects (PLINQ) to optimize queries for multi-core processors, and how to build custom parallel query operators
• Integrating the best LINQ to Objects patterns into your day-to-day coding
Projects
-
Agile Metrics Group
- Present
The Agile Metrics Group orchestrates, enables and performs useful research on effective product development. It seeks to increase the use of quantitative insight in product development, by making measurement more valuable to decision makers, more accessible to decision makers, more welcomed by participants, and more appropriately applied. We want to discover metrics that promote rapid adaptation to evidence. We seek to build a community collaborating on developing metrics and creating better…
The Agile Metrics Group orchestrates, enables and performs useful research on effective product development. It seeks to increase the use of quantitative insight in product development, by making measurement more valuable to decision makers, more accessible to decision makers, more welcomed by participants, and more appropriately applied. We want to discover metrics that promote rapid adaptation to evidence. We seek to build a community collaborating on developing metrics and creating better companies.
Other creators -
Honors & Awards
-
Brickell Key Award
Lean Systems Society
Award given for leadership in Lean software and systems community and discipline.
Recommendations received
2 people have recommended Troy
Join now to viewOther similar profiles
Explore more posts
-
Steve Jones, CFA
"I have only made this letter longer because I have not had the time to make it shorter." - Blaise Pascal'' This also applies to software demos. We've tightened the onboarding workflow and now have a 60 second demo of the journey from a blank slate in Claude Code to a deployed preview application on VeilStream.
1 Comment -
Sarah Hadley Hildebrandt
Agents just got a lot more real. Anthropic’s update gives Claude secure, remote access to tools like Linear—no local setup needed. This isn’t just about coding. It’s about context, continuity, and what it means to collaborate with AI. #AIagents #Claude #Anthropic #LLM #productthinking
-
Ari D. Bernstein
Excited to see this come together. For years, a lot of the most valuable E-Rate insights have been buried inside USAC datasets that are difficult to analyze in real time. Turning that data into usable intelligence can help vendors, consultants, and schools make much more informed decisions across the program.
-
Datastrato
"All-in-One" Suite 🆚 "Best-of-Breed" Approach Most organizations aren’t 100% on any single platform. Heterogeneity is the norm. That’s why metadata freedom is becoming a critical architectural discussion. 👉 Full read: https://lnkd.in/gHnrPvxd In this article, we break down: • why vendor lock-in matters more at the metadata layer • how multi-engine environments change governance • the tradeoffs between open catalogs and vertically integrated platforms This topic affects every engineering team operating across Spark, Trino, Flink, warehouses, vector stores, and AI systems.
-
Leanda Thomas
Addressing Fragmentation in the U.S. Cost Estimation Industry: The cost estimation industry in the U.S. is facing a pivotal challenge: fragmentation. Despite advances in digital tools and methodologies, more than half of U.S. teams report that their estimation workflows are still disconnected and siloed across departments. This fragmentation creates inefficiencies, from duplicated efforts to inconsistent data interpretation. The lack of interoperable systems means firms spend extra resources reconciling estimates and working through compatibility issues—ultimately impacting accuracy, project timelines, and stakeholder trust. In a sector where precision and agility are crucial for both competitiveness and client satisfaction, the persistence of disconnected workflows highlights an urgent need for integration, collaboration, and industry standards. By bridging these silos, we can unlock new levels of efficiency, innovation, and transparency—empowering firms to deliver better results and win greater confidence from clients. Are you seeing fragmented processes in your organization’s estimation practices? What steps have you taken to drive integration? I’d love to hear your thoughts!
-
ChenLi Wang
Always fun to host engineering dinners with Sumeet Vaidya and Crafting. Few takeaways for me: 1. We are living through one of the most dynamic periods in software engineering. Everyone by now knows to heavily leverage “Plan” mode upfront when coding with AI - but now we are hearing there’s even more time/tokens/turns spent on “Ask” mode before writing the Plan. And PRs of plans! We are literally flying up the abstractions curve. 2. Documentation used to be a luxury, but it is now cheap and easy. Documentation is increasingly agent-first, and markdowns in the repo seem to work a lot better than Confluence/Notion. 3. Historically the magic of GitHub was how it captured the collaboration context - PRs, discussions, reviews, commit history. Today so much of that context sits in prompt/chat history which does not go in GitHub or anywhere else! Feels like this needs to be solved. Openclaw’s memory system and Granola are both inspirations.
1 Comment -
Rose Hecksher Schamberger
Velocity and story points are capacity signals, not success metrics. Adoption is where value actually shows up (and it exposes the real bottlenecks: onboarding, UX, data quality, change management, and reliability). In my fractional work, I will often start by resetting the scorecard: - Adoption + retention signals (activation, usage depth, time-to-value) - Delivery health (lead time, failure rate, recovery time) - Learning loop speed (how fast we validate and iterate) Funny thing: when teams measure value, they stop shipping “more” and start shipping better and the roadmap gets easier to prioritize. #SaaS #EngineeringMeasurement #DeliveringValue
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content