Jaroslav Pantsjoha
Technical Director | AI Agentic Solutions Architect | GDE | Advisor
London, United Kingdom
Actions
I help enterprises move from AI MVPs to production-grade agentic systems - well-governed, secure, repeatably, and at scale.
Recognised Google Developer Expert and an avid builder. Enterprise architect and AI solution developer for my consulting business unit at Cognizant, where I lead Google Cloud agentic AI delivery. I spend much of my time thinking, unpacking and presenting my perspectives on what it actually takes to move POCs and agent solutions into production environments.
There's a lot of work happening behind the scenes for Q3/Q4 2026 — watch this space.
Badges
Area of Expertise
Topics
Harness Engineering: Building Your Lean, Mean Delivery Machine
Handing a developer a coding agent doesn't double your delivery. I've watched teams get busier, not faster.
What multiplies output is the harness around the model: the codified context, skills, hooks, sub-agents and ways of working, plus a squad built to run it.
One rule makes the whole thing work: if it's not in the repo, the agent doesn't know. And the part I genuinely didn't see coming: codifying the AI forced the humans to align.
I'll open a real harness I run on Google Cloud, delivering with a fraction of the planned team at a throughput that sounds made up until you see the headcount delta. I'll also be honest about the parts that still need tuning every single week.
The weekend pioneer who ships miracles is real.
But how many pioneers does your org have, and can you run an enterprise on pioneers alone?
This talk is how you get from one to many.
Takeaways:
1. The harness is what multiplies output, and the squad is what runs it. "It didn't happen overnight, and it didn't happen by accident."
2. You can't run an enterprise on pioneers. IC → codified harness → SME-led squad is how weekend-hero output scales without losing cadence or quality.
3. The harness can rot. Left uncurated it poisons its own context; the agent doesn't hallucinate, it just gets confused. Owning it is a weekly discipline.
Duration: 30 min (25 + 5 Q&A), scalable to 45. Audience: builders, architects, engineering leads. Deck built; drawn from live production delivery on Google Cloud. Part 1 of a Build / Secure / Scale signature set.
We Poisoned Our Own Context: The Human-Alignment Problem in Agentic Delivery
We obsess over whether the model will hallucinate. The failure I actually lived through was quieter and worse.
The model didn't hallucinate.
It got confused, because we had poisoned our own context.
Months of design docs, decisions and reference material piled up in one place, at different dates, some superseding others, none of it curated. The agents, and the people, started giving incoherent answers, and nobody could say why. In a team of five sharing one context bank, that's binary: you operate like an A-Team, or everyone inherits the same rot at once, multiplied.
This talk is the war story and the discipline that came out of it: context as a curated product with an owner and a retention policy, human-to-AI alignment as process we already know (vision / status / roadmap as the interface), and the human overlaps that AI amplifies instead of removing.
The alignment problem was never the model's. It was ours.
Takeaways:
1. Context poisoning ≠ hallucination. Stale, contradictory, uncurated context makes agents confused, and it fails silently until a human says "this doesn't feel right."
2. The same vision / status / roadmap interface that aligns humans aligns agents. You already know how to do this.
3. At team scale the blast radius multiplies. One shared context bank means one team's context debt hits everyone at once. Govern it like a product.
Duration: 30 min (25 + 5 Q&A). Audience: builders, architects, engineering leads. Part 2 of the Build / Secure / Scale signature set.
We Figured Out the Build Phase. Now, Let's Solve the Agent Run Phase
We've cracked the build. Anyone can stand up an impressive single-team agent demo. The run, the ownership, the adoption — that's where it's stalling, and good work is starting to fall by the wayside.
Solving for one business unit is a POC. A real organisation needs a governed, operational platform that lets many BUs run agents without each one reinventing identity, guardrails, ownership and rollback. Deployed and healthy is not the same as live and good.
Drawing on systems I run in production, this talk maps the run-phase operating model: the self-optimising flywheel we had to cage, the adversarial gate that reviews agent work before it ships, and human-in-the-loop approval as a first-class control.
Much of this maps directly onto native primitives in tools like Google Antigravity, so you build the governance, not the plumbing.
The one thing no IDE can decide for you: what counts as evidence in your domain.
Takeaways:
1. Build is solved; Run isn't. A single-BU demo doesn't survive contact with a multi-BU org. The gap is the operating model.
2. Cage the flywheel. Unsupervised self-optimisation quietly automates curve-fitting; supervised, gated, default-OFF is the honest posture.
3. Antigravity-native HITL gates, ask-by-default permissions, hooks and budget monitors give you the controls for free. You supply the domain evidence.
Duration: 30 min (25 + 5 Q&A). Audience: platform leads, architects, engineering leadership. Deck built; scheduled for delivery at SREday London Q3 (Sep 2026). Part 3 of the Build / Secure / Scale signature set.
Jaroslav Pantsjoha
Technical Director | AI Agentic Solutions Architect | GDE | Advisor
London, United Kingdom
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top