Anthropic’s 2026 report says developers use AI in 60% of their work but can fully delegate only 0-20% of tasks [1]. That’s a massive gap. I dug into the public research to understand why — and the answer isn’t what most people assume.
As someone who uses AI coding agents daily, I assumed the bottleneck was model capability. It’s not. The research points to something more fundamental.
There’s a growing gap in the AI agent conversation. Everyone’s talking about agents — autonomous systems that can reason, plan, and act. But most demos stop at “the agent wrote a nice email” or “it summarized a document.” The real challenge starts when you need an agent to interact with production backend systems, handle authentication, deal with partial failures, and return structured results that downstream systems can consume.
Over the past several months, I’ve been building exactly this: tool-equipped LLM agents operating in the payments domain. These agents can query transaction systems, look up order histories, check workflow statuses, retrieve standard operating procedures, and interact with ticketing systems — all through structured tool interfaces that the model invokes autonomously.
Here’s what I’ve learned.
My sister recently asked me for some help in running a comparison job between some large files spanning GBs worth of text. The most efficient solution I could find was Big Query.
So I tried some sample jobs by manually getting the data into a Google Compute Engine then to Google Storage and then loaded into a BigQuery dataset. The reason I had to use the Compute Engine was because this large file was available on a remote server with an ssh only access. Instead of asking my sister to install gsutil for uploading to Google Storage, I found it easier to just ask her to scp the file to a GCE.
Running and benchmarking python is easier if you have a standard place to run and share them. Kaggle fits this profile perfectly. It is a hosted service for your code which can be run and shared online.
https://www.kaggle.com/tusharm567
As an added advantage, I get direct access to loads of public datasets to play and sharpen my skills on.
from math import *
%timeit tan(atan(exp(log(sqrt(1*1)))))
This kind of stuff is helpful while learning and testing out algorithms and/or comparing different approaches to a solution based on compute.
Recently, I started working on some Raspberry Pi projects, but the most frustrating part of that experience was to access the raspberry pi during development.
I had the following options:
The second option is obviously the way to go, but this needs some refining.
After a long delay I finally mustered up the courage to build the query ranking module. Some scary stuff. Here are the problems I’ve been facing whenever I start building this.
So I am switching to a new approach of this as suggested by a colleague.
For now, I have chosen packet for the trial server even though the cost is high. The specs are good (even for entry level containers) and I got some credits to work with initially.
So I setup the ssh keys for login and got to work.
The github repository for parsey mcparseface and syntaxnet is available here and it also lists out the steps required to setup syntaxnet and get it working. They also provide a demo shell script file which accepts the input at stdin and passes it to a python script named parser_eval. The output which is an ascii tree of the pos tagged statement is posted to stdout.
Before I start building up the server for the application, I wanted to pen down the structure for the app. The structure is represented below.
Abbreviations:
Converting the Natural Language to Database Query project over to Javascript requires the implementation of the Web Speech API in a single Javascript file. I will discuss the steps taken for the integration in this blog post. Kinda unorthodox but here’s a reference before the content, if you want to experiment on Web Speech API using the official guide.
First we check if the API is available with the browser using the following code