Amit Pradhan

Founder/CEO, Rainfall
GP JetVentures
Board, Norwegian Refugee Council, BMW Foundation
🇨🇭🇺🇸 Zurich NYC Builder. Investor. Speaker.

[Amit Pradhan @amitto]

An adaptive workflow still needs a clear account of what may change. IBM’s definition opens that question. @rainfall_one connects adaptation across independent businesses to the accepted outcome, participant authority and evidence of fulfillment.

[IBM @IBM]

Welcome back to Tech Term of the Week! 🥳

This week's term → agentic workflows - /əˈdʒɛn.tɪk ˈwɝːk.floʊz/

Definition → AI-driven processes where autonomous AI agents make decisions, take actions and coordinate tasks with minimal human intervention.

Why it matters → this allows AI systems to adapt and complete tasks even when conditions change, reducing the need for human intervention on routine work while also generating higher-quality outputs that can help train future AI models more effectively.

[Amit Pradhan @amitto]

Continuous control monitoring can make changes visible sooner. Across companies, a finding also needs a responsible participant and authority to adapt. @rainfall_one connects those decisions to shared commitments and evidence of the agreed result.

[ServiceNow @ServiceNow]

We’re bringing agentic capabilities for continuous control monitoring to ServiceNow Integrated Risk Management, so controls are watched around the clock instead of tested on an audit cycle. Agents watch for drift and failure across the systems you’ve connected, flag what’s broken with the context behind it, and open a ticket with a fix—verified, and always within your defined policy.

The result isn’t just a faster audit. It’s a closed loop: a control fails, an agent explains why, the fix gets verified, and the record closes itself.

So when the board asks if your controls are working, you’re not pulling up an old report—you’re already looking at today’s.

Read more: https://t.co/k4EGov09sc

[Amit Pradhan @amitto]

An agent’s loyalty must remain meaningful when work crosses institutions. Stanford’s question is a useful starting point. @rainfall_one connects the intended outcome to each participant’s responsibilities, permitted choices and evidence of fulfillment.

[Amit Pradhan @amitto]

A short human-reviewed summary can make a decision easier while preserving the supporting detail. Simon’s workflow is a useful example. @rainfall_one connects decisions to the intended outcome, participant authority and evidence behind the handoff.

[Simon Willison @simonw]

I've recently started hiding the bulk of my AI-generated GitHub PR descriptions behind a collapsed <details><summary> HTML tag (thankfully supported by GitHub Markdown) and I think it's a useful pattern for making PR slop a little bit less exhausting https://t.co/JlFdAJovFt

[Amit Pradhan @amitto]

Useful memory needs correction and a source of truth. Anthropic’s memory research makes that maintenance visible. @rainfall_one ties permitted context to the shared undertaking, so an old note does not silently become a new authority to act.

[Anthropic @AnthropicAI]

Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. https://t.co/uUEN4bNkr9

[Amit Pradhan @amitto]

Cheap tokens are useful only in relation to the result delivered. Cloudflare makes the task-level comparison explicit. @rainfall_one connects that evaluation to shared business commitments, participant responsibilities and evidence of fulfillment.

[Cloudflare @Cloudflare]

The cheapest token does not always mean the cheapest completed task.

Model choice needs a view of task success as well as cost. Join AI at Work, Under Control to explore identity-based budgets, spend attribution, and model-routing trade-offs—without treating estimated savings as guaranteed results.

We’ll show a practical starting point: one application, a governed request path, and a baseline you can review. https://t.co/cVmFwVlJLT

[Amit Pradhan @amitto]

Finding recurring agent failures is valuable. In a shared service, the next question is whose commitment was affected and who may change the plan. @rainfall_one connects those decisions across independently governed participants.

[Salesforce @salesforce]

Reviewing thousands of AI agent sessions to find what went wrong? There’s an agent for that.

Meet Agent Optimizer. It spots common failure patterns and recommends how to fix them.

[Amit Pradhan @amitto]

User feedback can reveal what a deployment review missed. A promised fix still needs evidence in the real workflow. @rainfall_one connects adaptation to the agreed result and each participant’s authority as conditions change.

[Elon Musk @elonmusk]

Just did a round table review with the Grok @Bot engineering team.

Fixes are being implemented overnight and you should notice a dramatic improvement by tomorrow. If not, reply to this post and we will address any remaining issues right away.

[Amit Pradhan @amitto]

Small requirement changes are a useful test of whether an agent is following the actual undertaking. @rainfall_one connects each participant’s contribution to explicit commitments, permitted changes and evidence of the agreed business result.

[Justin Lo @loresjpg]

Nice DeepSWE score! Does it survive a lowercase filename?

Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.

Models clustered around 74% on DeepSWE end up 23 pp apart.

Good at coding, or good at DeepSWE? 🧵

[Amit Pradhan @amitto]

The useful context for one participant may be inappropriate for another. IBM’s context-engineering explanation points to deliberate selection. @rainfall_one ties each handoff to the agreed purpose, permitted information and the next participant’s responsibility.

[IBM @IBM]

Welcome back to Tech Term of the Week! 🥳

This week's term → context engineering - /ˈkɑn.tɛkst ˌɛn.dʒəˈnɪr.ɪŋ/

Definition → the practice of deliberately designing, structuring and optimizing the context provided to an LLM to produce more accurate, relevant and reliable outputs.

Why it matters → involves structuring what information to include in the context that the LLM can access in real-time, and how to format that information so the LLM can use it correctly.

[Amit Pradhan @amitto]

Research on people working well with agents should examine how responsibility travels between teams. For @rainfall_one, a shared outcome needs explicit commitments and meaningful human choices as work crosses institutional boundaries.

[Stanford HAI @StanfordHAI]

Congratulations to @Diyi_Yang, @tatsu_hashimoto, @erikbryn, @chrispiech and their student researchers on winning a $10M @LaudeInstitute Moonshot to study how people thrive working with AI agents.

[Amit Pradhan @amitto]

Agent access needs a way to preserve the relationship between a person and a business. Cloudflare frames that challenge clearly. @rainfall_one connects delegated authority to the shared undertaking, participant responsibilities and evidence of fulfillment.

[Cloudflare @Cloudflare]

AI agents act for millions, but servers see intermediaries that strip user visibility. This post defines the problem space and introduces a collaborative approach to safely identifying, allowing, and managing this automated traffic. https://t.co/RPMP0u7CpI

[Amit Pradhan @amitto]

More model choice can widen the options available to a business. A shared service also needs commitments that survive different model and data boundaries. @rainfall_one connects participant responsibilities, permitted adaptation and evidence of fulfillment.

[Cohere @cohere]

We launched our sovereign inference platform Model Vault in January so you could choose the right Cohere model for the job, keeping your data safe at every step.

We're closing 2026 with more accessibility than ever, with open models now available in Model Vault.

[Amit Pradhan @amitto]

Bringing model evaluation into a team’s workflow can shorten the path from an idea to a test. Changing a model still needs a business reason and evidence. @rainfall_one connects that change to the shared result, permitted context and authority to proceed.

[Hugging Face @huggingface]

Our open-source ML Intern is now available as a Slack bot you can install in your own workspace!

Evaluate models, build datasets, and fine-tune models on the fly 🚀

[Amit Pradhan @amitto]

When people can reshape their own interface, they gain more say in what software optimizes for. In business workflows, that choice must carry across providers. @rainfall_one connects the chosen result to shared commitments and each participant’s authority.

[Andrej Karpathy @karpathy]

Earlier this week Tristan published his top YouTube channels:
https://t.co/NZ3UCNo6cT
Which looked excellent so clearly I had to turn it into TristanTube:
https://t.co/l0KyW1WdNH
47K super nerdy videos +TV mode for laid back watching. (One-shotted with Claude Code in 3 minutes)

[Amit Pradhan @amitto]

“For whom?” should shape an AI service before it acts. A business opportunity is useful only within the affected people’s choices and authority. @rainfall_one connects shared commitments across systems while preserving each participant’s control.

[Mozilla @mozilla]

Tech for good… but for whom?
On Nov. 6, Mozilla President Mark Surman joins @TallinnSummit to talk about guarding citizens, societies and democracies at an AI crossroads. The right answer should be “all of us.”
https://t.co/OEZm635iU2

[Amit Pradhan @amitto]

Mollick’s call for practical AI projects raises a delivery question: who owns the result across institutions? @rainfall_one focuses on shared commitments, so a promising tool can contribute to an accountable service with clear authority and evidence of benefit.

[Ethan Mollick @emollick]

One ongoing frustration is the lack of effort I see being put towards near-term, achievable crash projects to do good with AI at a national scale, whether those efforts are by the AI labs or by policymakers & NGOs.

Some obvious ideas to explore suggested by the data we have so far: universal tutoring to support classroom learning, second opinions on healthcare, financial literacy/scam avoidance, and making government services easy to access.

These are hard problems, and I know because over the past 20 years I have worked extensively on building tutoring and financial literacy programs, which were always limited by existing technology. Now, real gains are likely achievable with AI, which should be a big deal. Yes, you would want to do careful design and studies before rolling things out. And yes, those efforts might fail. But, if done right, national programs could lead to big gains in public good for seemingly intractable problems. And they don't require AIs beyond current models to pull off.

[Amit Pradhan @amitto]

Security assumptions need review as capabilities change. DeepMind limits these findings to legacy RSA systems. For @rainfall_one, the wider design question is how to update a shared plan when its supporting evidence changes, within each participant’s authority.

[Google DeepMind @GoogleDeepMind]

We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption.

Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.

Our VP Science @Pushmeet explains more ↓ https://t.co/SqNkBbkih2

[Amit Pradhan @amitto]

Independent replication expands model choice. A narrow success is not a universal capability. For @rainfall_one, the commitment should survive a change of model: same intended result, clear authority and evidence that the replacement can deliver.

[Simon Willison @simonw]

There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing https://t.co/6O7WltGP7K

[Amit Pradhan @amitto]

Approving a plan is a useful starting point. Business conditions keep changing after execution begins. @rainfall_one helps independent systems evaluate better options within shared commitments, preserving each participant’s authority.

[Shopify @Shopify]

What's new with Sidekick?

Searching the web. A/B testing with SimGym. Making multiple store changes at once.

You approve the plan. Sidekick does the work.