Truly Agentic Means Reasoning, Not Rewritten Automation

Episode Summary

Ashish Bhatia argues that a lot of what the industry now calls agentic is the previous era’s automation rewritten in new frameworks. What makes a workload genuinely agentic, he says, is reasoning built in deeply enough to handle the corner cases that used to send engineers back to patch the rules, plus enough memory that the system does not suffer session amnesia between runs. From there he lays out three phases: humans assisted by AI, humans orchestrating agents they hand cognitive work to, and agents invited onto the team as coworkers in a hybrid workforce. Phase two, he says, is very much feasible today, and of phase three he points only at sparks. The line the rest of the conversation hangs on is his rule for getting there safely: before you own your AI, you should own your eval. He then applies it to model selection, arguing you start with the largest model and come down the capability curve.

Key takeaways

  • His opening claim is that the agentic migration is largely relabeling. The industry is transforming a lot of its automation workloads into agentic workloads, he says, and a lot of those are yesteryear’s automation rewritten in the new agentic frameworks

  • What makes a workload truly agentic, in his account, is where the thinking sits. You need reasoning built in deeply enough that the cognitive overload of building the automation moves off the human and onto the AI, he says, and that is what would make it truly agentic. The old way baked the rules into the system and let it run forever

  • The test he uses for whether reasoning is real is what happens at the edges. There are always corner cases nobody tackled, he says, which is why teams keep going back to fix their automations, and the promise of agentic reasoning is that those corner cases and outlier scenarios get handled gracefully instead. He calls that the key piece which is missing

  • The second thing he wants built in is memory, and he names the failure it prevents. Longer-term, memory and context retention matter so these systems do not suffer from session amnesia, he says, and can hold long-form context on what is going on and use it for the decision at the corner case, the way a person would

  • Phase one of his roadmap is assistance, and the label is Microsoft’s rather than his own coinage. Humans collaborate with AI agents and get augmented responses for whatever they are trying to get done, and he takes the phase one assistance name from Microsoft, where he had just come from

  • Phase two is where the human stops doing the work and starts dividing it. You offload some cognitive work to an agent while remaining the orchestrator, he says, breaking a large cognitive task into pieces and handing each to the right agent. His examples are deep research and code generation, and he says this is very much feasible today in a lot of companies

  • Phase three is the one he frames as a change of category rather than degree. You stop being an orchestrator of tools and invite agents onto the team, he says, moving from thinking about these things as tools to thinking about them as coworkers, digital employees, in what he calls a hybrid workforce where they do end-to-end work and call their own sub-tools

  • He does not claim phase three has arrived, but he points at the day’s news for the sparks of it. Newer models can use tools inside the reasoning workflow itself, he says, rather than reasoning, calling a tool, waiting for it to return and only then answering. Within the reasoning workflow they can call multiple tools and come back with an opinion

  • His model for building trust in an agent is the one every manager already has. You get a new employee, he says, and you treat an intern differently from a professional with five years of experience. The questions are the same either way: consistency of decision-making, consistency of outcomes, and trust built the way you would build it with a human colleague

  • The line the episode turns on is his rule about evaluation. Before you own your AI, you should own your eval, he says, and his reason is a question: if you do not own it, what do you own in the pipeline

  • Asked how you actually do an eval, he says the tool is beside the point. It is the determination of whether this is the cognitive task you want to offload and how you tell good outcomes from bad across the whole spectrum, he says, naming adversarial, wrong, hallucinated and inappropriate outputs as the classes to write down

  • Writing the eval down is what gives you somewhere to put a result. Being able to write down what constitutes good and not good for a given cognitive task gives you a baseline, he says, so that when an outcome is generated you can say which class it falls into

  • He runs the eval in two loops rather than one. In the dev or inner loop you tweak and improve until the agent reaches a threshold you are comfortable with, he says, and in the production or outer loop you monitor the runtime behavior in the wild, identify where things go off the rails, and feed those scenarios back into the dev loop before the next version ships

  • On model routing his answer starts with the consumer, and the reason is money. Users pick a favorite model and often reach for an expensive one to get a smallish task done, he says, so the serving companies can save a ton of dollars and tokens if they do the right routing and the right classification

  • For the enterprise he calls routing the same thing on steroids, and then names the cost. If you are not picking the model yourself it introduces another degree of variance, he says, where a model you pick is set in stone; a model picked on the fly becomes a new factor to validate, which is not necessarily bad but is another vector to check before you push

  • His worked example of routing already in production is Perplexity, and it doubles as the caution. You pick a tier there rather than a model, he says, and multiple models run under the hood before the answer reaches you, but those are Perplexity’s intentional choices. Hit a model router as your API layer and the variance sits in that layer, which is why he says you would want to own the classification yourself

  • His build-and-buy advice runs downward from the best model, not upward from the cheapest. Start with something off the shelf and pick the largest model you can to prove the goal is achievable, he says, then come down the capability curve model by model until you reach a breaking point, which is the most efficient price-performance model to run the scenario on

  • He connects that back to the eval, which is what makes the descent measurable. You can always pick the biggest and baddest model and not get value for your dollars, he says, so you run your evals across the models you are interested in, look at the outcomes, and let what you are paying in tokens decide

  • Asked what is over-hyped, he names prompt engineering and says the attention belongs one level up. We should think deeply about context engineering rather than just prompt engineering, he says, because models do not know your business or your data. Some of it is clever prompting, but a lot of it is bringing the right context

  • The leadership blind spot he names is building AI onto the old shape of the work. A lot of AI is duct tape AI, as he calls it, and rethinking AI-native use cases is what AI leaders need to think deeply about again

  • Where he expects the next transformation is vertical AI, and the reason is asymmetry. Models are becoming more open and more pervasive while data becomes more closed, he says, so enterprises sitting on large data have a unique ability to turn it into something valuable, which he expects to create a lot of new growth

  • His two leadership instructions are both about first-hand use. Be curious with AI and try the tools yourself, he says, because if you are not, somebody else is telling you what to do. And the fear of experimentation is something you can only tackle by experimenting more

About Ashish Bhatia

Ashish Bhatia is an AI product leader with more than twenty years in product management, across Microsoft and Nokia. At Microsoft he spent over a decade on AI platform work, helping bring generative AI to Power Platform and AI Builder and leading platform strategy for Copilot Studio, which is where the language of assistance, orchestration and hybrid workforce in this conversation comes from. He is now director of product management at Audible, part of Amazon, working on generative-AI-powered conversational discovery. He speaks and advises on agent platforms and responsible AI adoption. At the time of this recording he was leaving Microsoft and about to join Amazon, which is how he is introduced on air.

 

In this episode

00:42 Welcome
01:36 What the shift from automation to agentification means
01:42 Yesteryear’s automation rewritten in agentic frameworks
01:59 Reasoning built in deeply enough to offload the cognitive overload
02:30 Corner cases, and why teams keep going back to fix automations
02:54 Memory, context retention, and session amnesia
03:23 Is full reasoning with no human in the loop premature
04:13 Phase one: humans assisted, and augmented responses
04:39 Phase two: the human as orchestrator, offloading cognitive work
05:19 Phase three: agents invited onto the team as coworkers
06:08 The announcement that day: tools used inside the reasoning workflow
06:54 What breakthroughs would make full autonomy safe
07:20 Onboarding an agent the way you onboard a new employee
07:35 Consistency of decision-making, and building trust
07:58 Before you own your AI, you should own your eval
08:26 The tool is a means: deciding what good and bad look like
09:03 Writing it down gives you a baseline to classify against
09:12 The dev loop and the production loop
10:01 What invisible model routing unlocks for the enterprise
10:42 Routing by the outcome, the question and the intent
11:07 Saving dollars and tokens with the right classification
11:26 For enterprise it is the same thing, on steroids
12:09 An aside on enterprises already running more than one model
12:33 Perplexity: you pick a tier, not a model
13:46 How this changes the build and buy decision
13:52 Start with the largest model and prove the goal is reachable
14:37 Running the evals across candidate models, and what the tokens cost
15:06 The lightning round
15:21 Over-hyped: prompt engineering, versus context engineering
15:42 Bullish for 2026: document understanding and processing
15:55 The blind spot: duct tape AI, and AI-native thinking
16:26 Vertical AI, and enterprises sitting on closed data
16:55 Asked where to follow him, and the answer the transcript loses
17:05 His defining edge: absorb as much as you can
17:38 Be curious, because otherwise someone tells you what to do
18:15 Wrap-up

In Ashish’s words

“Before you own your AI, you should own your eval”

— Ashish Bhatia   (07:58)

“A lot of those are yesteryears automation rewritten in the new agentic frameworks, if you will.”

— Ashish Bhatia   (01:42)

“we are moving away from thinking about these things as tools to coworkers, employees, digital employees of the world”

— Ashish Bhatia   (05:19)

“these things don’t suffer from kind of session amnesia as much, and have long-form context and long context of what’s going on.”

— Ashish Bhatia   (02:54)

“how do we understand consistency of decision-making, consistency of outcomes, and then start to build trust with them”

— Ashish Bhatia   (07:35)

“In the end, most users just care about the outcome. They don’t care about picking the model.”

— Ashish Bhatia   (11:26)

“You can always pick the biggest and the baddest model, but you might not get the value for your dollars”

— Ashish Bhatia   (14:37)

“A lot of AI is duct tape AI, as I call it.”

— Ashish Bhatia   (15:55)

“Be curious with AI. You have to try these tools to be able to realize their benefit, because if you’re not, somebody else is telling you what to do”

— Ashish Bhatia   (17:38)

 

Resources

  • Ashish Bhatia on LinkedIn: Where he posts, and the best place to follow his work on agent platforms and responsible AI adoption

  • Audible: Where he is director of product management, working on generative-AI-powered conversational discovery. He was about to join at the time of this conversation

The announcement he points to

  • Introducing GPT-5: Announced on 7 August 2025, the day this conversation was recorded, and the release he is pointing at when he says at 06:08 that there are sparks of phase three already. It is a unified system whose reasoning model uses tools as part of its thinking rather than reasoning first and calling a tool afterwards, which is the change he describes. He does not name it on air

Ideas and terms discussed

  • Yesteryear’s automation rewritten: His diagnosis, and the argument the rest of the conversation rests on. Organizations are moving automation workloads onto agentic frameworks, and a lot of what results is the old thing in new clothes. The difference he insists on is where the thinking happens: rules baked in by an engineer and left to run, versus reasoning deep enough that the system handles the case nobody wrote a rule for

  • Corner cases: His test for whether an agentic rewrite is real. Traditional automation sends its builders back repeatedly to patch the scenarios they did not anticipate. What he wants from agentic reasoning is those outliers handled gracefully at runtime instead, and he calls that the key piece which is missing rather than a thing already delivered

  • Session amnesia: His name for what memory is supposed to fix. Without retained context an agent starts each run from nothing, so the long-form context that a person would carry into a judgment call is unavailable exactly where the corner case needs it

  • The three phases: His roadmap, and the spine of the episode. Phase one is assistance, humans working with AI and getting augmented responses. Phase two is the human as orchestrator, breaking a large cognitive task apart and handing pieces to the right agent, which he says is feasible in a lot of companies today. Phase three is the hybrid workforce, where agents are invited onto the team as coworkers doing end-to-end work and calling their own sub-tools

  • Own your eval: The rule everything else in the conversation is measured against, and his own formulation. The eval is the written definition of what good and not good look like for a specific cognitive task, across the whole spectrum from adversarial to hallucinated to merely inappropriate. Written down, it becomes a baseline that any output can be classified against, which is what makes trust something you can build rather than assert

  • The dev loop and the production loop: How he wants the eval used, in two places rather than one. In the dev or inner loop you tweak until the agent clears a threshold you are comfortable with. In the production or outer loop you monitor runtime behavior in the wild, catch what is going off the rails, and route those scenarios back into the dev loop before the next version ships

  • Invisible model routing: A model picked for you, by intent, instead of by you. On the consumer side he thinks it saves the serving companies real money because users pick expensive models for small tasks. In the enterprise it is the same thing on steroids, with a cost: a model you choose is set in stone, and a model chosen on the fly is a new degree of variance to validate. His conclusion is that you would want to own the routing classification yourself

  • Down the capability curve: His method for choosing a model. Start with the largest model off the shelf and prove the goal is achievable at all, then step down model by model until you reach a breaking point, which is the most efficient price-performance model to run the scenario on. The evals are what make each step measurable, and what the tokens cost decides the rest

  • Context engineering: What he thinks the industry should be attending to instead of prompt engineering, which is his pick for the most over-hyped thing in generative AI. Models do not know your business or your data, so some of the work is clever prompting and a lot of it is assembling the right context to hand them

  • Duct tape AI: His phrase, offered without a definition, for the AI he thinks is added to work that was designed without it. He names rethinking AI-native use cases as the leadership blind spot holding back returns, and duct tape AI is the counterpart he wants replaced, which is the same argument one level up from context engineering

  • Vertical AI: Where he expects the next transformation, on an asymmetry he describes: models are becoming more open and more pervasive at the same time as data becomes more closed. Enterprises sitting on large data therefore hold something that is getting scarcer rather than cheaper

Named on air

  • Microsoft: Where he had just come from, named at 04:13 as the source of the phase one assistance language he uses for his own roadmap. He spent over a decade there on AI platform work

  • Cursor and Replit: His examples at 04:58 of phase two in production for code, alongside deep research from Google as the example for research

  • Perplexity: His worked example at 12:33 of model routing already shipped: you pick a tier rather than a model, several models run under the hood before the answer reaches you, and the choices are the vendor’s rather than yours

  • ChatGPT, o3, GPT-4.1 and GPT-4.5: The models he names at 10:26 and 10:42 when describing how a consumer picks by task, reaching for a reasoning model for research and a different one for creative writing. GPT-4.5 had been withdrawn from the API a few weeks before this conversation but was still in ChatGPT, which is where he says it is still available

 

Frequently Asked Questions

 
 
 
 
 
 
 
 
 
 
 
Previous
Previous

An AI Committee Needs Every Department and Real Authority

Next
Next

Detect Intent, Then Tailor Every Screen to the Person