The decisions behind dependable AI
What we delegate, who remains accountable, and how we judge the outcome.
Model Fieldnotes — by Swapnil Upganlawar
From AI pilots
to dependable work.
I help engineering and product leaders turn AI pilots into dependable work, with clear ownership and costs they can defend.
My view: Judge AI by the work it completes, the responsibility people retain, and the full cost of getting there.
By Swapnil Upganlawar Building with AI. Sharing what holds up.
The fieldnotes method
Start here
Read the argument. Inspect the evidence.
A perspective, a teaching lab, and a test you can run yourself.
Watch an agent fail and recover
Inspect a lost response, apply one repair, and replay the same situation. No model or API key.
Run the retry experiment
100 intentions. 20 lost responses. Compare the service ledgers and take the test with you.
Latest in AI
Release analysisJev and Strands Decider: where decision models fit
A hosted service or an open model? A close look at Jev and Strands Decider, the confidence scores that do not quite match, and the economics of putting a small model in front of a larger one.
Philadelphia · 3 October 2026
From the conversation / Wharton
AI adoption starts with the customer problem.
I joined the technology panel at WEMBA’s 50th Anniversary & Reunion in Philadelphia on October 3. The reunion brought together more than 700 Wharton alumni from around the world.
My reflections on AI adoption, data readiness, leadership, governance, trust, and the cost of getting useful work done.
Read my panel reflections01 / My perspective
Where I think AI is heading.
Three shifts shaping the decisions I explore here.
From answers to outcomes
Completed work, errors, retries, and human intervention belong in the same evaluation.
For leadersMeasure cost per successful task and decide which failures are acceptable.
Explore the workThe cost of a completed taskFrom demos to dependable systems
Reliability depends on the surrounding software, controls, and verification as well as the model.
For leadersDesign permissions, recovery, and evaluation into the product from the start.
Explore the workAgent Reliability FieldbookFrom model selection to organizational design
Delegating work to AI changes workflows, ownership, and escalation.
For leadersDecide what to delegate, who remains accountable, and when people intervene.
Explore the workLeadership & OrganizationsThese are personal perspectives. The Fieldbook’s simulations explain specific mechanisms; they do not establish these forecasts.
Leadership & Organizations
How AI changes the work of leading.
Management, team design, architecture, and accountability. My analysis, with readings that make the evidence and its limits visible.
Explore Leadership & Organizations02 / The open lab
Ideas you can get your hands on.
New release / Open source / v0.1.0
Agent Failure Recipes
Run a failure. Leave with a test.
Duplicate actions, premature completion, and uncontrolled retries. Three executable TypeScript recipes let you inspect the failure, apply a repair, and take the regression test into your own code.
Controlled fixtures, not model benchmarks. MIT licensed. No API keys.
Deterministic simulationsPublic beta
Agent Explainer
See how AI agents work.
And why they fail.
Watch a failure. Inspect the cause. Apply one repair and replay. Three browser-based simulations make the invisible parts of an agent system visible.
Read the Agent Reliability Fieldbook for each repair’s mechanism and production limits.
No API keysOpen sourceRuns locally in your browser
agent-explainer / duplicate-action
From everyday life
Personal projects
Personal project / Built for my son
Veer Reading Quest
Read. Picture it. Connect the clues. Prove your answer.
I built this for my son, Veer: short stories, room to think and draw, and a habit of finding evidence in the passage. Two learning levels, a calm pace, and saved work for parents to review.
Explore Veer Reading QuestThe design question
How do you help a child slow down, understand a story, and explain why an answer fits?
20 original stories · Core missions need no AI03 / Perspectives
Ideas for the decisions ahead.
The AI score went up. What actually improved?
Four October studies show why a better AI score isn't always a reason to launch. Here's what I'd check about the measurement first.
Essay7 min read
By Swapnil Upganlawar
100 operations. 20 lost responses. What did the retries create?
A reproducible fault-injection experiment: both callers acknowledged 100 operations, but ordinary retries left 120 writes. Stable operation keys kept the ledger at 100.
Essay6 min read
By Swapnil Upganlawar
The cheapest AI attempt is not always the cheapest useful result
Compare the cost of acceptable outcomes, including failed attempts, retries, tools, and human review, with the task cost calculator.
Fieldnote1 min read
By Swapnil Upganlawar
AI adoption starts with the customer problem
Reflections from the technology panel at WEMBA’s 50th anniversary: working backwards from customers, getting the data right, and making adoption a leadership responsibility.
Fieldnote4 min read
By Swapnil Upganlawar
04 / The reading room
Five foundations for understanding AI agents
One paper in depth. Four more in perspective. What the evidence says, where it stops, and why it matters for builders.
Read the collection- 01↗
ReAct
Connecting a plan to observations from the world.
2022 · In depth - 02↗
Toolformer
Learning when an external tool is useful.
2023 · Research note - 03↗
Lost in the Middle
Having context and using it are different abilities.
2023 · Research note - 04↗
Reflexion
Carrying feedback into the next attempt.
2023 · Research note - 05↗
τ-bench
Checking whether an agent reliably did the right thing.
2024 · Research note
Developments / Beyond the headline
What’s changing in AI.
New releases, the evidence behind them, and the decisions they could change.
Jev and Strands Decider: where decision models fit
A hosted service or an open model? A close look at Jev and Strands Decider, the confidence scores that do not quite match, and the economics of putting a small model in front of a larger one.
Source-based analysis · Not independently tested
Announced
Analysis updated
Release analysis12 min read
By Swapnil Upganlawar
Model Fieldnotes · Automated briefing · Sep 28, 2026
Hugging Face expands Apple local AI work ↗
Hugging Face announced GGUF inference support in Transformers for Apple Silicon and hired oMLX creator Jun Kim, framing both moves as investments in local AI tooling.
05 / Behind the fieldnotes
Hi, I’m Swapnil Upganlawar.
My background spans more than two decades: machine learning research at Carnegie Mellon, founding software ventures, and AI leadership at Gartner, Amazon, and AWS. I hold an MBA from Wharton and an MS in Bioinformatics from Carnegie Mellon. Model Fieldnotes brings that experience to practical decisions for engineering leaders and senior builders.
Contributor, Forbes Technology Council.
Keep a thread of curiosity open
Follow the next fieldnote.
Follow signed perspectives, guides, and research in your feed reader. Automated briefings are available in the combined feed.