Praveen Seshadri

@JamesonCamp Have you seen that there's already 1001 tabs in Salesforce and it needs a training guide to even find something there?

Salesforce is gonna ship everything your favorite AI CRM startup built as a tab in a dropdown menu So will Adobe. So will Intuit. And everyone’s gonna forget those startups existed 30 days later

@mcuban Yes spot on. And since agents typically need to touch content and take actions in multiple SaaS apps, no one SaaS app is going to be suitable as "the agent platform"/ orchestration layer. Value shifts away from SaaS to the agent platform. ai.plainenglish.io/the-fading-of-…

If true and agents work on top of enterprise software, doesn't this eliminate the need for per seat pricing by the software companies ? The coin of the realm for agents and AI in general is tokens. I don't see how enterprise software reconciles this conflict. Particularly when the agent "shops" for the most cost effective path with in an enterprise. I think the enterprise software companies will be able to charge for creating and managing agents and how they engage for companies that can't. But I don't see how the revenues stay where they are. Thoughts ?

@buccocapital That's an optimistic scenario -- infinite market opportunity, so free up people to move to more opportunity. Or, the market opportunity is limited so it's a race to cut costs and take market from competitors. And that is what the financial markets seem to be assuming

One easy way to not be a doomer about AI is to go talk to people running actual businesses Simple example: Let’s say you use AI to answer your tier 1 support cases. Many are doing this for 30-60% of cases depending on business complexity What is actually happening is that these businesses are now doing two things: 1. Training the tier 1 reps to be tier 2 reps, that are doing more service, maybe selling, or doing more hybrid work with product to triage things that broke through AI because they are either bugs or feature requests 2. Taking the cost savings and hiring more CSM’s who can do concierge-type service or drive upsell/focus more on Churn If your business is firing on cylinders, you just move the people/hiring to the next blockage point or opportunity to drive revenue

@jeff_weinstein @stripe The API usage volume for well-built enterprise agents isn't much (it is comparable with the human-scale workload it is replacing). The only obstacle has been the reliability of the agents. But that's also being figured out: thunk.ai/ai-at-work/the…

a SaaS company reached out to us after OpenClaw-style agents are sending ~50k requests per hour (wayyy above normal site usage). the company now wants to charge agents small amount for reads and writes with @stripe machine payments. if this is also you (or may soon be), say hi.

@nasdaily The vibe-coding stuff isn't the issue. AI agents automating the workflows is the issue. The only obstacle has been the reliability of the agents. But that's also being figured out: thunk.ai/ai-at-work/the…

Just spent $1m buying SaaS stocks because: 1) markets are being irrational 2) the "vibe code everything" yourself is hype. 3) valuable companies are heavily discounted 4) people underestimate distribution.

@ctindale The vibe-coding alternative makes no sense. But AI agents automating the workflows is the issue. The only obstacle has been the reliability of the agents. But that's also being proven out: thunk.ai/ai-at-work/the…

@emollick Now that LLM benchmarking is reaching a point of diminishing returns, we've started doing application/workflow benchmarking. It's a better test of the "scaffolding". Would be interested in your thoughts: thunk.ai/ai-at-work/the…

Many benchmarks use LLMs as a judge of correctness, typically a smaller, cheaper model. This paper shows weaker judges are not able to evaluate smarter models. A benchmark is really a triplet of dataset, model, judge & judges are increasingly the bottleneck being saturated.

@philotheis49253 @stevesi Totally agree this is critical for reliable (enterprise) AI automation. But also, imo, this is not the job of the LLM, it is the job of the scaffolding/platform that harnesses the LLM. Eg: thunk.ai/ai-at-work/the…

@pravse @stevesi One big challenge is just getting the AI to understand when it should escalate in the first place

There is little doubt agents (aka functions, subroutines, programs, scripts, stored procedures, ) will be deployed at massive scale for everything from buying something, writing something, drawing something, inquiring about something, answering something, and so on). The question is not whether code can be writtern (or prompted, spoken, etc.) to do one of those tasks. The real question is what will the agent do when something unintended or unexpected happens? The history of software automating tasks is not about getting the expected case to work, it is about handling the unexpected, “exception handling”. Any look at a human process today involves two things: • software to handle the expected • a human to navigate and tell the software what to do after undertaking the effort to solve for the unexpected and undocumented exception Any time spent at an airline counter, post office, DMV, Comcast office, and more will immediately tell you the problem almost always boils down to an exception to the process. The human is “making” the system do things it should not do. Often this means overriding the security of the system and taking responsibility for the result. Example. Comcast just recently automatically gave me a $100 credit (for down service) but it never showed up on my bill. I spent a half hour with the bot trying to talk it into the credit with no luck. I called up customer service and the person saw the credit but there was no software path to produce the credit. It took that person a day offline working the system to make the credit show up on the bill. The “agent” could award the credit but the billing and customer service systems could not follow through. Every thing we can think of an agent doing requires us to think through every possible path to handle an exception. This has been the history of automating processes with software and is why automation is difficult. …rdcoresoftware.learningbyshipping.com/?utm_source=na…

@fchollet The value today is in the enterprise workflows have to get done. Humans do the workflows (biggest cost), they record their progress in SaaS apps (secondary cost). When agents do the workflow, human cost shrinks and the SaaS app is just a schematized db with low (human-scale) TPS

The maximalist form of my thesis is basically this: SaaS is not about code, it is about solving a problem customers have and selling them the solution. Services + sales. If the cost of code goes to *zero*, SaaS will *not* go away. It will *benefit*, since code is a cost center.

@levie Where is the app and the user? Where is the report glowing? Where are the tickets and the workflow rules flowing? They have passed like an outdated style, like a free trial, The usage has shifted down and to the right, away from our sight, and into AI. How did it come to this?

Today, software is primarily built for people to use (directly or indirectly). But it's very clear that there will be trillions of agents in the future, executing every type of task for us imaginable.  Agents will be deployed for coding, processing loans, reviewing insurance claims, executing financial transactions, acting as personal assistants, and every other known task in the economy. As a result, we're going to see a shift in who we have to increasingly build tools for.  So many new opportunities are rapidly emerging right for building for agents. Agents are going to need seamless identities across platforms. They're going to need file systems and databases to store off their work, sessions, and important data they're sharing. They're going to need tools for collaborating with people. They're going to need safe ways of spending or managing money. They're going to need computers to execute code and other tasks in. And so on. In many cases, the tools and systems that the human users are already working with will be the natural tools for these agents to leverage. There are many areas where the highways have already been built, and agents will ride right on top of those. In other cases, there will need to be new capabilities that emerge due to the scale and change in use-case that agents represent. In either case, these tools need to be API-first, as agents will leverage these tools like a developer or machine would have previously. CLIs/APIs are their native tongue.  The complex part is that building for agents introduces new challenges vs. building for people. They require far more oversight than people do, and they don't get the same right to privacy as people. They can't be held responsible for the work that they're doing, but rather the person that launches them into their task must be (for now). They don't quite know when they've run astray and can't execute the task at hand. These are just a small set of things that become the new complexities that need to be anticipated when building for agents. We’re entering a completing new era of software development and infrastructure that will be built out. Wild times ahead.

@fchollet Where is the app and the user? Where is the report glowing? Where are the tickets and the workflow rules flowing? They have passed like an outdated style, like a free trial, The usage has shifted down and to the right, away from our sight, and into AI. How did it come to this?

Folks at Anthropic can correct me if I'm wrong, but I believe Anthropic uses Slack, Zoom, Figma, Notion, Workday, and Google Workspace. Correct?

@stevesi As an aside, I don't believe "code will be written" is the best framing of the problem. That suggests that deterministic code is the materialized form factor for workflow intent. I think most work agents will directly interpret intent specified in natural language.

@emollick On the other hand, there are other agentic harnesses like Thunk.AI specialized for greater process specificity and AI reliability (versus the generic broad "think"/"try" Claude harness where it's ok to get it somewhat wrong)

All those products where building an "AI agent" meant defining a series of basic prompts linked together deterministically through a flowchart with separate RAG inputs are looking pretty dated right about now (yes, that is basically every agent product released in 2025)

@emollick So the comparison should be between the Claude harness and products like n8n (deterministic workflow harness+ AI steps). n8n is definitely not AI agents.

@emollick There seems to be a basic misunderstanding. Claude code spits out "code" which is deterministic software. The value of Claude Code (or CoWork) itself is that it is an agentic harness.

@stevekrouse What's "a ton" of PDFs? 10,000? 1000? 100,000? Small or large PDFs? Might have a good solution to recommend but it depends on this..

I have a customer with a ton of PDFs they want an LLM on top of, but we're hitting context window limits Is there a high-level API that lets me upload a bunch of PDFs, and then provides a "tool" that I can give to an LLM?

@emollick Seems like you could have called it "Free Weights" and it would have also served a dual purpose 😃

The hardcover book of GPT-1’s weights that Claude Code designed, produced, and sold (including the cool cover which visualizes the numbers in the volume) actually came in the mail today and it looks really nice. I never touched any code or did any design or any API to make this.

@emollick @emollick that is a strange definition. If you went back five years and asked that question (what's "insane" to expect AI to do), you'd be badly wrong. So why do you feel it would be right now? The "training set" for our expectations is backward looking

It really isn’t that hard to see the jagged frontier of AI. Just think about the parts of your job that are vital but that you would be insane to expect an AI to do, even if agents get 10x better. Thats the frontier The more you use AI the more accurate those assessments will be

@asaio87 I agree. The threat to SaaS isn't vibe-coding (though that will drive competition). The threat is that agents run the workflows medium.com/ai-in-plain-en…

Companies will not vibe code their own internal tools to avoid spending $50 on a subscription. Development, security and maintenance of such tools takes time and effort, thus money, Headaches that companies will not want to assume. So people who say saas apps is dead due to vibe coding, are 100% wrong.