If artificial intelligence is creating new technology at breakneck speed, it is also creating a new vocabulary just as quickly. Terms such as “prompt engineering” entered the mainstream a couple of years back, while newer phrases like “loop engineering” and “forward deployed engineer” are beginning to appear across job listings, startup pitches and industry conversations.
Are these terms reflecting genuine shifts in how AI systems are being built and deployed? What exactly do these new roles and concepts mean, and why are they attracting so much attention?
In this week’s edition of Tech Talk:
- Meta gets India cred with Kunal Shah as WhatsApp boss
- Big Tech is panicking about AI and terrifying workers
- AI Tool of the Week: ChatGPT Scheduled Tasks
Anthropic’s Head of Claude Code, Boris Cherny. (AP)
Anthropic PBC co-founder and Claude Code creator Boris Cherny, who had previously proclaimed that “software engineering is dead”, now asserts that the era of manual AI prompting is drawing to a close.
- Prompt engineering focuses on crafting the right instructions to get the best possible response from an AI model.
- Loop engineering is about designing workflows in which AI systems repeatedly reason, act, check their work and improve their outputs.
For example, a prompt engineer might write a prompt asking an AI to summarise a legal document. A loop engineer would build a system that reads the document, identifies ambiguities, searches relevant regulations, reviews its own conclusions and flags uncertain cases for human review. As AI becomes more agentic, the challenge shifts from writing better prompts to designing better processes, making loop engineering a natural evolution of the role.
Loop engineering is about encoding expertise. The best loop engineers are often not the people who know the most about AI. They are the people who understand a domain such as media, law, medicine, finance, or software development, and can translate that expertise into a repeatable sequence of checks, decisions and actions.
- Prompt engineering: You ask a question. The AI model answers.
- Context engineering: You provide the right documents, examples, memories, and tools for the model to deliver a better outcome.
- Loop engineering: You define how the model plans, acts, checks its work, uses tools, recovers from errors, and decides when the job is done.
These stages coexist. A coding agent still relies on good prompts. A research assistant still benefits from careful context management. Loop engineering simply becomes more important as you move from “generate an answer” to “complete an objective”.
This video by Anthropic engineers is instructive. DeepLearning.AI founder and Coursera co-founder Andrew Ng, says: “100% of my tasks are now done by AI agents—hype has exceeded my expectations. Loops is the next step. In 3-6 months, everyone will be using self-improving loops. No more prompting.”
That said, loop engineering can be significantly more expensive than prompt engineering because it typically requires multiple model interactions rather than a single request and response. A simple prompt-based task might involve one model call consuming a few thousand tokens. A loop-based system, however, may ask the model to plan, retrieve information, evaluate its findings, revise its output, verify its conclusions and, in some cases, seek human review before producing a final answer.
Each of these steps consumes additional tokens—meaning, a task that might require 5,000 tokens as a prompt could easily consume 25,000–100,000 when implemented as a loop. The higher cost comes from repeated reasoning, the accumulation of context across multiple steps and the need to process information returned by tools such as search engines, databases or code interpreters.
The trade-off is improved reliability
A single prompt may be sufficient to summarise a document, but a loop can check for errors, verify facts and revisit uncertain conclusions before responding.
As a result, loops are particularly valuable for high-stakes applications like software development, legal analysis and financial decision-making, where accuracy matters more than efficiency. Yet, as models become cheaper, token costs are likely to fall faster than labour costs. The cost of AI inference has declined dramatically while model capabilities have improved. If that trend continues, the economic penalty for using loops will become less significant over time.
Forward Deployed Engineer too is not an entirely new term, but AI has turned it from a niche role into one of the hottest jobs in tech. The term became widely associated with Palantir Technologies Inc., where engineers were sent directly to customer sites to understand real-world problems and build solutions alongside users rather than from a distant headquarters. Today, AI companies such as OpenAI, Anthropic, and Cursor have embraced similar roles.
An FDE is a hybrid of software engineer, product manager, consultant and AI workflow designer. Unlike a traditional software engineer deployed onsite by an IT services firm—whose role is usually to implement predefined requirements, integrate systems or maintain software—an FDE is expected to identify the problem, redesign workflows and determine how AI can create measurable business value. The focus is less on coding features and more on deploying intelligence.
As AI shifts attention from prompts to agents and business processes, FDEs are increasingly designing Human-AI workflows, evaluation loops and organisational change, making them an AI-native evolution of customer-facing engineering roles.
Meta can buy talent, but success takes a lot more
Lately, Meta Platforms Inc. CEO Mark Zuckerberg has been reaching for his cheque book ever so often to solve strategic problems.
Confronted with mounting competition in AI, for instance, Meta spent billions on talent. It recruited high-profile researchers, founders and entire teams from startups such as Dreamer, Maltbook and Scale AI (whose co-founder Alexandr Wang was asked to lead its quest for super-intelligence).
Now, Zuckerberg appears to have taken a similar playbook to WhatsApp. Meta has invested $900 million in Indian fintech startup Cred for a minority stake and hired its founder Kunal Shah to lead the chat platform globally.
Kunal Shah, co-founder at Cred and new global head of WhatsApp.
India is central to Meta’s growth story. Of WhatsApp’s 3.3 billion monthly active users, a large chunk are Indians. Yet, for all its popularity, WhatsApp has been unable to grab ad revenue and has struggled to evolve into the financial powerhouse that Meta had envisioned it to be. Enter Kunal Shah—a serial entrepreneur with success in payments, credit and other financial services. Clearly, Meta wants to monetise WhatsApp quickly.
Recruiting startup founders can infuse fresh ideas and entrepreneurial energy, but that in itself does not create successful services. In other words, although Meta has shown that it can attract top talent, how it translates this into sustainable business growth is the actual test it faces. Here’s the full article.
AI TOOL OF THE WEEK
By AI&Beyond, with Jaspreet Bindra and Anuj Magazine
The AI capability unlocked today is ChatGPT Scheduled Tasks
What problem does it solve? Most professionals quietly run the same set of tasks every week: the Monday competitor scan, the Friday project update, the weekly inbox check. Each is a 20-minute ritual of opening tabs, reading, and summarising that requires no real thinking. Just showing up.
ChatGPT Scheduled Tasks closes that loop. Describe the task once, set the cadence, and ChatGPT runs it: searches the web, checks your connected apps, writes the output, and delivers it on schedule.
A new dedicated Scheduled page in the sidebar lets you view, pause, edit, or delete every active task in one place.
How to access: https://chatgpt.com
ChatGPT Scheduled Tasks can help you:
- Kill the loop: Stop doing the same search-summarise-share sequence each week and let ChatGPT run it on a cadence you control.
- Brief yourself automatically: Wake up to a fresh AI-written summary of competitors, news, or updates pulled from your own connected apps, delivered before your first meeting.
- Monitor two layers: Set ChatGPT to watch the web for signals and check connected apps like Gmail or Google Drive, alerting you only when something changes.
Example: A senior manager sets up two scheduled tasks. Here is how each plays out:
Web monitoring: Competitive intelligence
- Set the task: Ask ChatGPT “Every Monday at 7:30 am, search for news about [competitor] from the past seven days and give me the top five developments.”
- Add a filter: Refine with “only send if there is something material, otherwise skip” to cut noise on quiet weeks.
Connected apps: Inbox triage
- Set the task: Ask ChatGPT “Every Friday at 4:00 pm, check my Gmail for unresolved client threads this week and summarise what needs a reply.”
- Get the brief: ChatGPT scans your connected Gmail, identifies open threads, and delivers a prioritised list—no manual inbox-diving required.
An Edge case worth knowing: Write operations in connected apps, such as drafting replies or creating documents in Google Drive, are currently limited to Enterprise users.
What makes ChatGPT Scheduled Tasks special?
- Monitoring, not just pinging: ChatGPT checks the web, your connected apps and writes a fresh output each time, not a static calendar alert.
- Plain language, no setup: Describe the task in one sentence and it runs indefinitely with no Zapier flows, scripts, or technical configuration.
- A category eliminated: This removes an entire class of recurring low-value work from your week, across both external signals and your own inbox and files.
Note: The tools and analysis featured in this section demonstrated clear value based on our internal testing. Our recommendations are entirely independent and not influenced by the tool creators.
Most AI-led mainframe migrations will fall short: Gartner
AI is not a silver bullet for mainframe migration, and many companies may be better off modernising existing systems than replacing them, suggests a new Gartner report. More than 70% of mainframe migration projects launched in 2026 will fail to deliver their expected benefits because companies are overestimating what Generative AI can do, according to the research firm.
Gartner cautions that many organisations are being swayed by vendor claims that AI can quickly and easily modernise decades-old systems. In reality, migrating complex legacy applications remains difficult, risky and expensive.
ALSO READ | The Sword of Damocles that’s Anthropic Claude
As Claude Mythos falls, OpenAI GPT-5.5 Cyber rises
Last month, OpenAI granted the US government early access to GPT-5.5 for national security testing and evaluation. The latest frontier model is designed to tackle complex real-world tasks. On Monday, OpenAI expanded access to a cyber-focused version, GPT-5.5 Cyber, through a limited release to what it calls “trusted defenders”.
To that end, the company unveiled Daybreak, a platform that combines its frontier cyber models, Trusted Access for Cyber, Codex Security workflows, and third-party partners. The goal is to help approved defenders identify vulnerabilities, prioritise risks, generate fixes, and produce evidence within existing security and software-development workflows.
ALSO READ | Who gets to use Frontier AI?
Big Tech is panicking about AI and terrifying workers
A three-decade veteran of Microsoft Corp. is pushing back against what he describes as a culture of manufactured alarm inside the technology industry. Brad Smith, who serves as vice-chairman and president of the company, says influential voices in tech have been consistently wrong about the pace of disruption, consistently hypocritical about the solutions they propose, and consistently careless about the human cost of the panic they generate.
ALSO READ | We can’t let AI giants eat the economy: Nadella
MORE ALSO READS
Tech Talk is a weekly newsletter by Leslie D’Monte on everything happening in the world of technology and AI. Want this newsletter delivered to your inbox? Subscribe here.
