“AI Co-founders” with Chris Hefley & Troy Magennis

Hey all! We’re back from the summer break with Season 2 of Hard Boiled Software. This season’s theme is “people in the AI neighborhood" in which we seek out unique roles adapting to new technology. Today we have a story from two AI co-founders working across two startups. Double the fun!

It was such a pleasure having an in-depth conversation with Chris and Troy. These guys are fascinating, pushing two AI-first companies forward at the same time. There’s a lot in here for anyone working in product discovery or innovation.

The Guests

Chris Hefley is co-founder and CEO of Predictability at Scale, the company behind UsageTap, and co-founder of FocusedFit, an AI-native executive recruiting platform. He previously co-founded and led LeanKit, the enterprise Kanban platform that was acquired by Planview in 2017 and is now known as Planview AgilePlace. A developer by training and a product leader by practice, Chris has spent his career building software companies and obsessing over how to hire the people who make them work.

Troy Magennis is co-founder and CTO of Predictability at Scale and the founder of Focused Objective, where he became one of the best-known voices on probabilistic forecasting and Monte Carlo simulation in software delivery. He is the author of Forecasting and Simulating Software Development Projects, a Brickell Key Award recipient, and a former VP of Development and Global Architecture at Sabre. These days he builds open-source tooling for teams trying to bring some rigor to AI-assisted engineering.

Summary

Chris Hefley and Troy Magennis met in the middle of a conference hall in Seattle, with Troy holding a laptop and teaching Chris Monte Carlo simulation off an Excel spreadsheet for forty minutes. That conversation turned into years of collaboration at LeanKit, and eventually into a pair of AI companies they now run together.

In this episode, Dave talks with both of them about what happens when the constraint that shaped software development for decades suddenly moves. When coding is no longer the slow step, uncertainty doesn’t disappear. It relocates, landing on strategy, product decisions, go-to-market, and the operating costs nobody budgeted for. Troy puts it plainly: at 600 miles an hour, being one degree off your strategy is a very big miss.

The conversation covers their honest pivot story (including the part where a very good product arrived too late for its market), why software features alone no longer hold value, what AI inference is doing to SaaS margins, and how UsageTap and FocusedFit are each betting on a different kind of attention. Along the way, Chris explains why he doesn’t believe in A players, Troy shares his favorite forecasting tell, and Chris walks through how he runs a nightly council of AI agents that wake him up to a dozen pull requests every morning.

Links & Resources

Guest Links

* Predictability at Scale: Chris and Troy’s company

* UsageTap: usage metering, limits, forecasting, and billing for AI-native apps

* FocusedFit: AI-native recruiting platform with agents that run the full search

* Alvarez Search: healthcare executive search firm where FocusedFit took shape

* Troy Magennis on LinkedIn

* Chris Hefley on LinkedIn

* Troy’s Substack

* Focused Objective: Troy’s forecasting and metrics consultancy

* Focused Objective Resources on GitHub: Troy’s free forecasting spreadsheets and presentations

* Forecasting and Simulating Software Development Projects by Troy

Open-source and AI Engineering Tools

* PromptOpsKit: prompts and system instructions as versioned, linted, compiled application assets

* CapabilityKit: capability-as-code, giving coding agents a dependency tree of what the system does

* DeliveryTower: agentic engineering performance tracking, including capturing “regrets” as future constraints

Books & Articles Mentioned

* The Mom Test by Rob Fitzpatrick: how to get honest customer feedback when people want to be nice to you

* Who: The A Method for Hiring by Geoff Smart & Randy Street: the A-player hiring book Chris references

* Netflix Culture Deck: the first edition of the “we’re a team, not a family” professional sports team framing

* “Big Ball of Mud” by Brian Foote & Joseph Yoder

Tools, Frameworks & Concepts

* Monte Carlo Simulation: probabilistic forecasting that rolls in the uncertainty of past outcomes

* LeanKit / Planview AgilePlace: the enterprise Kanban tool Chris co-founded

* OpenRouter: AI model gateway, recently acquired by Stripe

* Stripe: payments, and UsageTap’s billing integration

* Hugging Face: open model hub, recently acquired by Nvidia

* Nylas: email and calendar connector API Chris pays for rather than maintaining himself

* Lovable: AI app builder, cited for treating free-trial inference as marketing spend

* Product Hunt: the launch platform where vibe-coded products meet their first inference bill

* Myers-Briggs and DISC: the personality instruments Chris compares to “horoscopes for men”

People Referenced

* Jim Benson & Tonianne DeMaria: organizers of the Seattle Lean Kanban open space event where Chris and Troy met

* Sanchit Maruti, MD: psychiatrist, UVM Larner College of Medicine associate professor, and FocusedFit co-founder

* W. Edwards Deming: “A bad system will beat a good person every time”

* Brian Foote & Joseph Yoder: authors of “Big Ball of Mud”

Key Points

Monte Carlo and Where Uncertainty Lives Now

Troy picked up Monte Carlo in the defense industry in the 1990s, where it was simply how risk was managed on fixed-price work. He brought it to software because a hope-and-prayers approach to on-time delivery was never good enough. AI hasn’t made forecasting obsolete. It has pushed us to confront where uncertainty lives.

Key points:

* Handoff and coding uncertainty is shrinking, but “are we building the right thing” and “does it stay working in production” are as uncertain as ever

* Historical data only helps if your way of working hasn’t changed, which is exactly the trap for vendors who forecast from years-old data

* Chris’s revelation in that first conversation: you don’t need thousands of data points to forecast usefully, just a small, recent sample from your new way of working

* Troy kept his forecasting tools in spreadsheets on purpose, so people had to engage with the assumptions

The Constraints Have Moved

With coding no longer the bottleneck, every other constraint in the system suddenly matters. Roles that used to have months of breathing room between quarterly planning cycles are now fielding “it’s done, what’s next?” within days.

Key points:

* Product, marketing, sales, training, and strategy can no longer hide behind a slow development constraint

* Troy’s warning: at six miles an hour, being one degree off strategy leaves time to react; at 600, it’s a big miss

* Teams can now wander into competitive markets they never should have entered, because poorly considered features ship faster than ever

* Chris argues the Lean and Agile principles still hold, but every ceremony from big room planning to the retrospective deserves fresh eyes

Pivoting Honestly

Chris and Troy trace the path from BlockedApp, a tool for surfacing blockers across large enterprises, through an AI gateway product that ran into OpenRouter, to the cost and usage features customers actually wanted. That extraction became UsageTap, and an example app built for an executive recruiting firm became FocusedFit.

Key points:

* Their blocker, dependency, and team forecasting work was genuinely good, and genuinely too late for a market that no longer plans the way it did in 2025

* Troy on accepting “the death of some of your time”: no amount of wanting it changes what the market needs now

* Chris’s discovery toolkit is mostly conversations, including treating early investor pitches as free consulting, while watching out for the Mom Test problem

* Building in public matters more than ever, since traditional marketing can’t cut through the inbox noise

* Showing up with a working product is now the baseline, and investors and accelerators increasingly expect around $10K in monthly revenue before taking the meeting

Where Software Value Lives Now

If anyone can build the features, the features aren’t the value. Chris argues that what customers pay for is attention: someone else keeping up with the compliance, legal, and ecosystem changes they don’t want to track forever.

Key points:

* Chris pays Nylas for email connectors he could build in days, because maintaining them for years means chasing compliance and data tenancy rules

* In recruiting, anyone can vibe code a candidate sourcer, but AI hiring law is a fast-moving target that can get you in real trouble

* The 80% gross margin SaaS model is under pressure; Chris expects healthy companies to land closer to 55 to 60%

* Free trials used to cost almost nothing to offer; AI is claiming marketing and go-to-market budget, and Chris now models $9 to $15 of inference exposure per trial

* Recent deals like Stripe acquiring OpenRouter and Nvidia acquiring Hugging Face raise real questions about how the economics of AI infrastructure will settle

UsageTap and the Grown-Up Economics of AI

UsageTap taps off the usage data from AI calls and turns it into plans, limits, forecasts, anomaly alerts, and billing. For Troy, it’s an operations tool in a market flooded with AI development tools and starved of AI operations tools.

Key points:

* One feature vibe coded on a Tuesday night pushed Chris’s dev account from about $5 a day to $25 to $30, and UsageTap’s anomaly detection caught it

* Launch day on Product Hunt can mean a five-figure inference bill for a product that never budgeted for it

* Chris’s caching lesson: putting the current timestamp at the top of a system instruction meant nothing ever got cached

* Troy’s view of FinOps: track cost per task, not tokens, and match the right model and thinking level to the job

* Both see open-weight models taking over small and medium tasks, shifting costs from tokens to hosting

* Troy expects future apps to be called mostly by agents and APIs, which opens a billing gap few products have solved

FocusedFit and the Myth of the “A-Player”

Chris’s obsession with hiring goes back to LeanKit, where he learned the hard way what it’s like to hire for roles he’d never done. FocusedFit runs an expert recruiter’s workflow as a team of AI agents, with behavioral science built in by co-founder Dr. Sanchit Maruti, and deliberately looks for alignment rather than personality.

Key points:

* FocusedFit looks for misalignment between four things: the role, the candidate, the hiring team, and market realities

* A CTO in San Francisco for under $200,000 is a role-to-market misalignment; a candidate who thrives on chaos joining a compliance culture is a conversation worth having

* A hybrid “FocusedFit Live” model pairs self-service with AI-fluent partner recruiters, and the context they build stays with the company when they leave

* Speed matters: long interview processes lose good candidates to other offers

* Chris doesn’t believe in A players; performance depends on the system, the timing, and the person’s life, echoing Deming’s “a bad system will beat a good person every time”

* Troy’s forecasting tell: the teams with the recognized A players and 10x developers are the least predictable, because everyone else keeps borrowing them

Running a Team of Agents, and Coding Without Regrets

The episode closes with how both guests actually build with AI. Chris runs FocusedFit with a tiered set of agents that meet in a nightly council, while Troy has turned his own painful lessons into open-source tooling.

Key points:

* Chris’s setup is simple: a repo of agents, a chief of staff agent and a chief engineer agent on top, and GitHub Issues as the durable queue for agents and humans alike

* Agents operate within guardrails, stop when they need him, and Chris adjusts those guardrails as he unblocks them, so the same question doesn’t come back

* A compliance agent flagged a “rejected” column in the candidate pipeline; the fix was an auditable record of who moved a candidate, when, and why

* Troy’s PromptOpsKit treats prompts as code, CapabilityKit gives agents a dependency map of what a change might break, and DeliveryTower helps teams turn regrets into constraints

* Troy’s advice for AI builders: when something goes wrong, build something that reduces that pain next time

* Dave’s closing reminder: it’s still software, and a big ball of mud just arrives a lot faster

Hard Boiled Software is hosted by Dave Laribee.

This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit newsletter.nerdnoir.com