Day One at Global Agile Summit 2026: Thirteen Talks
Notes from a remote attendee. What each speaker brought, and what stuck.
The summit ran fully online this year. Four parallel tracks at most slots, two keynotes bookending the day. Watching from Bulgaria meant I caught everything in catch-up rather than choosing in real time — which has its own benefits, since I could rewind talks and skip ones that didn’t earn the attention.
What surprised me by the end of the day: how much the five tracks (Keynotes, AI in Organizations, People, Agile in Gaming, Agile in Construction) kept arguing the same things from different angles. None of the arguments were new. The convergence was — speakers working in their own corners for decades, landing in roughly the same place at roughly the same time.
Going through the talks in air-time order.
Marcus Bullock — Building Human-Centered Tech in an Unpredictable Economy
The opening keynote. Bullock spent eight years in prison, came out to forty-one job rejections, and built FlickShop — a postcard app modelled on the postcards his mother sent him while he was inside. His framing: human connection is infrastructure, not a feature. The product metric he optimizes for is decreased time in the app. Faster send, more postcards, more connection.
Emotionalize the problem before productizing it. My Tier-3 investigation skill and code-review orchestration both started from the same place — not “what’s the generic best practice” but “what is the specific friction I deal with, and how would I want a system to ease it.” Bullock framed this as designing with people, not for them. Same idea I’ve been running on without naming it.
He mentioned in passing that he’s building agents and Claude orchestration. Didn’t open that thread. The talk’s emotional weight is in the prison-postcards-FlickShop arc, and that’s the right weight.
Brenda Valencia — From 2 to 9 Projects: Scaling Precon with Scrum
Construction track. Brenda’s preconstruction team went from drowning at two projects to running nine, by introducing standups, retros, and weekly Tuesday cross-training rotations. Concrete byproducts: near-zero attrition, every team member promoted within two years, time to film an office-parody video and still hit deadlines.
The story I want to remember is the resistance pattern. The chief estimator told her standups were a waste of time; he’d rather work. Six months in, he was the one defending the ceremonies because he’d seen what they unblocked. Brenda’s line — “intentional time talking to each other ends up giving you more time” — is the cleanest formulation of why lightweight rituals beat raw individual focus over a long period.
The takeaway that landed for me: when capacity expands, the question is what you do with the excess. Greedy answer is “do twice the work.” Brenda’s team chose stability and skill compounding instead. Same trade-off shows up in any team that makes itself faster — including a team where the speedup came from AI orchestration rather than from process.
Kari Koivistoinen — Building Worlds, Not Burnout
Gaming track. Twenty-plus years as an executive producer, including Alan Wake and Quantum Break at Remedy. His thesis: crunch is not a project problem, it’s a clarity problem. Indecision and late trade-offs train teams into reactive cycles; the survival response gets mistaken for heroism.
The example that stuck: Kari joined a burnt-out cinematics team mid-project and opened with “cinematics is new to me, teach me.” Not strategic vulnerability — actual vulnerability, used to build the relationship before he needed it for hard conversations. Six months later the team was raising risks early instead of swallowing them, and the producer’s job had quietly shifted from triage to upstream protection.
For me, this maps cleanly onto mentoring. The juniors I have worked with over the years opened up with risks and confusions only after the relationship was established for its own sake. Trust is not a deliverable.
Kari also said something I have been trying to articulate for years: “sustainable delivery is not about working less, but deciding better.” I will be quoting this.
Dave Westgarth — Vibe UX: Prototyping at the Speed of Thought
AI track. Westgarth’s claim: prototyping has collapsed from a two-week sprint to minutes, and the differentiator is no longer delivery speed — it is taste. Concrete tools mentioned: Lovable, v0, Claude Code, Replit, Base 44, embedding prototypes in Miro as iframes for parallel feedback, instrumenting prototypes with telemetry that an LLM then interprets.
This was the most directly applicable talk of the day for the work my support team does on the marketing website. We watch newly-registered users get stuck in the same places, week after week. The version of that loop most teams settle into is slow — observation gets logged, prioritization runs at its own cadence, friction survives several rounds of reports. Westgarth’s loop is: observe friction → generate three candidate fixes in Lovable → put them in Miro → cohort-test → ship the strongest. The inversion is generation cost. Once that drops near zero, which variant to pursue becomes the actual job.
Westgarth’s other line worth preserving: don’t write perfect prompts yourself. Rough-sketch the idea, ask Claude to format it for the prototype tool, feed Claude’s output downstream. I do this already for code; I had not extended it to UI prototypes.
Michael Dougherty & Pete Oliver-Krueger — Shift: To Usability Theater
People track, co-presented. The framing — “usability theater” — is not the negative version. It’s a deliberate practice: record real users, then screen the recordings with the team like you would screen a film. Make the user’s struggle a shared experience inside the room.
Three test types they distinguish: scripted (near release, verify against spec), closed (mid-cycle, give a goal and watch silently), open (early exploration, no script, catch the unknowns). The enforcing rule across all three: silence. The biggest failure mode in user testing is the observer talking — explaining the task, asking leading questions, narrating what the user must mean. Open follow-ups only: “Tell me more. What are you thinking?” Then nothing.
This describes my team’s weekly session-recording reviews almost exactly, with one gap I will fix: we summarize what we observed rather than letting people watch the clip. Findings without the video get rationalized away. The video does not.
Jeff Patton — Why We Build the Wrong Things
People track. Patton wrote User Story Mapping. The talk was a long argument that the wrong thing gets built when teams optimize for output velocity and feature count instead of for what real users can actually use. AI, in his framing, amplifies the problem rather than solving it: a large company released an AI-redesigned version of its product with the same features but worse usability, and the market responded predictably.
The story I will remember: Patton early-career, told users were too busy to talk to. He went down to the trading floor anyway to watch quietly. Within minutes the traders turned around, asked what he was doing, and started volunteering everything they wished was different. The wisdom: people want to help if they believe you’ll listen. Earlier teams had burnt the channel by promising attention and then ignoring it.
His closing line: “no ideas survive the first customer exposure.” I have watched this be true for years and never had a sentence for it.
Anna Lavrova — When Leaders Say “People First” but the System Says Otherwise
People track. Lavrova works as an organizational coach, with sixteen years of agile-transformation work and educational programs co-built with Ukraine’s Ministry of Digital Transformation. Her diagnostic frame is Albert Hirschman’s Exit, Voice, and Loyalty — Hirschman’s 1970 book on how people respond when an organization or service fails them.
Four responses to systemic dysfunction: loyalty (stay silent and endure), exit (leave), voice (speak up), neglect (disengage cynically). Lavrova’s argument: in a world where exit is harder (AI-driven layoffs, market uncertainty), loyalty becomes the dominant response, and organizations reward the silence — which reinforces the dysfunction. Voice should be rewarded. It rarely is.
Her sharpest test for whether an organization actually puts people first: not the rhetoric, but the weak signals. Her example was a CTO habitually late to meetings while leadership talks about respecting people’s time. The system is teaching what it actually values; nobody has to read the values document to learn it.
I will be reading Hirschman this month. The framework is too useful to leave at the level of one summit talk.
Shawn Wallack — Even With AI, Your System Will Never Be Better Than Its People
AI track. Wallack’s argument lands in the same place as Patton’s, from the engineering side: AI compresses execution but does not change the constraint. Bottlenecks shift; they do not disappear. The new bottleneck is judgment — review, approval, accountability — and those still belong to humans.
His sharp point about career pipelines: as junior-level execution shrinks, the entry ramp narrows. The orgs eventually run out of people who know how to direct and review the systems they depend on. Anyone running a long mentoring practice should be tracking this carefully. I am.
AI as the chainsaw, not the colleague. The lumberjack still decides where to cut, why, and who is accountable for the tree falling correctly.
Kat Antonowicz — Beyond Frameworks: The Human Side of Agile
Gaming track. Antonowicz came into game production from social work — burnout-and-depression focus, specifically. Her central observation: passionate creative professionals burn out differently from other workers because they treat the work as a calling. Boundaries dissolve from the inside.
Her case study was an estimate-resistance project. Senior team, 20+ years in-house, no agile background. Resistance to estimates was not stubbornness — it was job-security anxiety dressed as methodological objection. The fix took months: reframe estimates as protection against late-cycle pressure, not measurement; pair adoption with a manager the team already trusted; normalize estimation failures as learning rather than evidence of incompetence.
The transferable skill she names is from social work: stay curious instead of judgmental. “There is almost never resistance just to show resistance. There is always an underlying reason.” I will remember this on the next round of process change I have to push through.
Greg Rog — Building an AI-First Organization
AI track. Rog (Grzegorz Róg) is co-founder and CTO of Easy Tools. The talk was the densest of the day on actual orchestration architecture. He builds named, persona-based assistants — Lou the project manager, May the marketer, Steve the mentor — each preloaded with company context, then layers reusable skills on top of them. Multi-assistant workflows pass output between assistants in one thread: Lou writes the project summary, May reads it and drafts the LinkedIn post.
The pattern that lands hardest: process beats outcome generation. Rog refuses one-shot prompts. He writes the first version of a UI tooltip himself, asks an LLM to extract the style and pattern, builds a skill from the extracted template, then automates all future tooltips through that skill. The skill is the product. The first artifact was the spec.
I built a multi-agent classifier-tagger-translator-quality-loop pipeline for my big cats news project in early 2025, before “agent orchestration” was a phrase anyone was using. Watching Rog’s talk was the closest I have come to seeing my own working pattern named externally. The same context-preloading, the same skill stacking, the same insistence that the AI does not generate from scratch — it executes a pattern the human committed to first.
His distinction between assistant and agent is worth keeping: assistant follows a defined skill, agent decides which tool and which output. Most of what I have built and most of what is useful at production scale is assistants, not agents.
Umar Ijaz — Independent Game Development: The Natural Home for Agile
Gaming track. Four-person studio (Agapo Games), one shipped game across a six-month timeline, two more in prototype. They use a three-column Kanban board, WhatsApp internally, and direct conversation with players via Discord, Steam forums, YouTube, Facebook, email. “We don’t use the word scrum, Kanban, or anything.”
The example I will not forget: an achievement-hunter player who got frustrated, downgraded the Steam review, then got pulled into the team’s Discord through a direct conversation. He is now a regular QA-feedback partner. A 70-year-old player buys the game on every platform and emails feature wishes. The entire studio’s external feedback loop is built out of relationships, not analytics dashboards.
Ijaz’s hard limit: this works up to about eight people. Past that, agility-by-conversation breaks and you need structure. “If scaling means we become double-A or triple-A, it is a totally different ballgame.” He did not romanticize the small-team life. He named it as one of several legitimate states.
For my own personal-project work — the big cats news pipeline, the Telegram-to-multi-platform thread orchestrator I have been building — this is the natural register. One person, conversation with the data, no ceremony. The pattern fits the scale, and stretching it past its scale would just rebuild the bureaucracy I am avoiding.
Felipe Engineer-Manriquez — From Chatbot to Companion: AI That Remembers Your Jobsite
Construction track. Felipe’s argument: a dumber AI with persistent memory of your project beats a smarter stateless one. He runs a memory database with ten thousand-plus knowledge items — thirty years of research papers, project specs, personal metadata — searchable semantically. He rotates between Claude Opus 4.6 (1M token window), Qwen 3.5 (running locally), Google Gemma (also local), depending on the task. The database survives the rotation; that is the point.
Two technical details I want to keep. First, model rotation via markdown export and reimport — “give me everything about me” — copy-paste into the next model, reload context. This is the most direct answer to vendor lock-in I have seen articulated. Second, the cron-scheduled retrospective: Felipe ran 24-hour AI sprints with no review for thirty days, automated a weekly retro via cron, and his AI named Osito promoted itself to scrum master when it noticed the gap. The feedback loop closed itself.
This is exactly the architectural problem I solved differently in my big cats news pipeline. I have flip-flop detection through three layers — full-article hashing, per-sentence hashing, correction-pair detection — and my equivalent of memory rotation is handled at the database level rather than markdown export. Felipe’s framing made me reconsider whether the markdown-export pattern is more robust to model change than my per-step persistence. Worth testing on the next pipeline pass.
His provocation: “GPT is smart, but it is like a kid coloring with crayons. I am gonna show you how to use spray paint.” Memory is the spray paint.
Clinton Keith — Game Development, as We Know It, Is Ending, and It Could Get a Lot Better
Closing keynote. Keith was the CTO at High Moon Studios when scrum first came into game development in 2003. He has spent the last twenty years coaching studios. His thesis: AAA game development is structurally locked into a $500M-budget, hit-or-miss model that wastes roughly 50% of a project’s cost on coordination overhead, and AI-plus-better-tools makes a smaller, faster studio model viable in a way it was not a decade ago. He does not predict AAA’s death. He predicts AAA’s stagnation, while teams of twelve with the right tooling move past them.
The story he opened with — the lowest-rated programmer on his JIRA dashboard, who turned out to be the highest-leverage engineer on the team because he was helping animators with custom tools instead of closing tickets — is the sharpest argument I have heard for why ticket-throughput metrics rot a craft. Measure the wrong thing and the team optimizes for it. The team that is measured against fun produces fun. The team that is measured against tickets produces tickets.
Red work is repetitive, measurable, automatable. Blue work is creative, subjective, irreducible to a metric. AI handles red work; humans do blue work. The orchestration I have built — code-review skill across fifteen-plus agents, Tier-3 ticket investigation across three-plus agents, QA test-plan agent — is the red-work-to-AI handoff Keith is recommending, built from the support-engineering side instead of the games-industry side. Same shape, different domain.
Threads across the day
Three things kept recurring across tracks, and I want to write them down before the day blurs:
Judgment is not displaced; it is relocated. Wallack, Patton, Greg Rog, Felipe Engineer, Clinton Keith all made versions of this argument. The shape is the same: AI compresses the execution layer, the bottleneck moves to review and decision, the role of the human is to direct and to be accountable. None of them framed AI as a replacement. All of them framed it as an amplifier of what was already in the room — including the dysfunction.
Watching beats theorizing. Patton on stock traders. Dougherty and Oliver-Krueger on usability theater. Bullock on building with the people closest to the problem. The pattern repeats so consistently across the day that it stops feeling like advice and starts feeling like a missing default.
Clarity is the burnout-prevention layer. Kari Koivistoinen on indecision and late trade-offs as the actual cause of crunch. Lavrova on systems that reward loyalty-as-silence and call it commitment. Antonowicz on resistance as fear-with-an-underlying-reason. The thread is structural — not “support people more,” but “decide things earlier and with better signal so support does not have to be heroic.”
I will be reading Hirschman, adding a cron-scheduled retrospective to my big cats news pipeline the way Felipe did, and trying Westgarth’s Lovable-into-Miro loop on at least one of next month’s marketing-site iterations. Day Two is tomorrow. I will see if the convergence holds.