Can Harness Engineering Give Kashmir’s Tech Startups a Fair Shot at the AI Race?

by Dr Nawab John Dar

A new engineering discipline called harness engineering is changing how the world builds reliable artificial intelligence systems. Unlike earlier shifts in the AI race, this one rewards patient engineering discipline over access to giant compute budgets. That has real implications for India’s own growing AI ambitions, and more specifically, for the technology entrepreneurs of Jammu and Kashmir, provided some real and documented gaps are addressed honestly rather than talked around.

In just a few years, the language used to describe how engineers build with artificial intelligence has shifted through several distinct emphases. Each shift sounds, from the outside, like marketing. It is not. Each one marks a genuine and traceable change in where the hardest, most expensive problem in AI actually sits. Understanding that change, and understanding why the latest term was borrowed from horse tack, called a harness, tells us something real about where opportunity for the next generation of technology entrepreneurs will come from, in India broadly, and in Jammu and Kashmir specifically.

The most recent shift traces to February 5, 2026, when Mitchell Hashimoto, the co-founder of HashiCorp and the creator of the widely used infrastructure tools Terraform and Vagrant, published a long personal blog post titled My AI Adoption Journey. In a section of that post titled “Engineer the Harness,” he laid out a simple idea he did not think was new enough to deserve a fresh name, though the industry gave it one anyway.

His idea was this: whenever an AI agent makes a mistake, do not just correct it and move on. Take the time to change the environment around the agent so that the same mistake becomes structurally impossible to repeat.¹

Six days later, on February 11, 2026, OpenAI published its own account of an internal experiment in which a small team built and shipped an internal beta product containing roughly one million lines of code over five months, with zero lines written by hand and about 1,500 pull requests along the way, all generated by its Codex agents while humans steered.²

The post used the same term in its title, Harness Engineering: Leveraging Codex in an Agent-First World. The phrase spread quickly from there. LangChain and Datadog subsequently used the same language in describing their own agent development practices, and by the middle of 2026 harness engineering had become a widely used term for an emerging approach to building more reliable AI agents.

Three engineering emphases, three bottlenecks

To understand why this third emphasis matters, it helps to walk through what came before it, because each one addressed a real bottleneck of its time. One useful way to understand this evolution is through three overlapping engineering emphases: prompt engineering, context engineering, and, increasingly, harness engineering. Each emerged in response to a different practical bottleneck.

The first widely discussed emphasis was prompt engineering, as large language models became capable of performing very different tasks depending on how instructions were formulated. In the years after OpenAI’s GPT-3 demonstrated that a single model could perform wildly different tasks depending on how a request was worded, the model’s raw ability was the limiting factor, so practitioners spent hours testing small variations of wording to coax a better answer out of the same model. Adding a phrase like “let us think step by step” or opening with “you are an expert in this field” could measurably change output quality. It was real, useful work, but it was fragile, because improvements from one task rarely carried over cleanly to the next.

A second emphasis, increasingly described as context engineering, became prominent in the middle of 2025. On June 19, 2025, Tobi Lütke, the chief executive of Shopify, posted publicly that he preferred a different term to describe the skill his engineers actually needed, describing it as the art of supplying an AI model with all the information required to plausibly solve a task.³ Six days later, Andrej Karpathy, the AI researcher and former Tesla director of AI, endorsed and sharpened the idea, describing it as the delicate art and science of filling a model’s limited working memory with exactly the right information for the step at hand.⁴

The distinction mattered because increasingly capable models made the surrounding information, tools, instructions and state available to the model a much more important part of system design. The real work had shifted to curation, deciding which documents, which prior conversations, which tool results, and which instructions actually belonged inside the model’s attention at any given moment. Anthropic’s applied AI team formalised this thinking in an engineering post published on September 29, 2025, titled “Effective Context Engineering for AI Agents,” which described the discipline as curating and maintaining the optimal set of information available to a model during a task.⁵

Around the same time, a technical report from Chroma tested eighteen frontier language models on straightforward retrieval tasks and found that every one of them lost accuracy as context length increased, often well before the maximum advertised context window would have suggested that performance should become problematic, a pattern the researchers called context rot.⁶ Bigger context windows alone did not solve the problem. Someone still had to decide what belonged inside the window.

The third emphasis, harness engineering, is the one we are living through now, and it exists because context engineering, even done well, still left a gap. An AI agent could have exactly the right information in front of it and still take the wrong action, forget a rule it had already been given, mark a task as finished when it was not, or repeat an error a human had already corrected once. The problem was no longer only about what the model could see. It was about what surrounded the model: what tools it was allowed to use, what checks ran before and after each action, what happened when it made a mistake, and whether that mistake could ever recur. This is the problem harness engineering addresses directly.

What a harness actually is

One useful way to think about the relationship is: Agent = Model + Harness. The model provides the underlying capabilities; the harness supplies the surrounding tools, instructions, environment, checks and feedback mechanisms that determine how those capabilities are used, and that make an agent’s behaviour more constrained, observable, testable and repeatable in practice.⁷

It helps to borrow the industry’s own metaphor for describing this idea, because it is a genuinely useful one outside the software world: a harness, as in horse tack. The metaphor is useful because a harness does not make a horse more powerful. It gives that power structure and direction. In much the same way, harness engineering does not primarily mean making the underlying model more capable. It means designing the environment in which an agent works so that its capabilities can be directed, observed, tested and corrected.

In practical terms, a harness can involve several interacting components. One component decides exactly which tools an agent is allowed to use at a given moment, and what data it may see, so that an agent handling, for instance, a hospital patient’s records is not accidentally given access to an unrelated system. Another is a verification loop, a set of automated checks that run after every action the agent takes, so that a claim of task completion is tested rather than taken on faith. A third governs memory, deciding what the agent should remember from earlier in a long task and what it should be made to forget so it does not get confused.

A fourth is made up of guardrails, hard limits that stop an agent from taking an action it should never be allowed to take regardless of what it reasons its way into. A fifth is observability, the ability for a human supervisor to see exactly what the agent did and why, after the fact, which becomes essential the moment something goes wrong. There is no universally agreed list of harness components; different teams emphasise different combinations of tools, context, verification, state management, observability and constraints. None of this is exotic technology. Much of it is the same discipline of testing, logging, and access control that experienced software engineers have practised for decades, applied now to a new and less predictable kind of co-worker.

Anthropic’s own engineering team described a concrete version of this problem in a report published on November 26, 2025, before the term harness engineering existed but describing exactly the practice it names. Their engineers found that a strong AI coding model, left to work across many separate sessions on a large software project, would either try to do everything at once and run out of working memory halfway through a feature, or would look at a partly finished project and wrongly decide the job was done.⁸

Their fix was not a smarter model. It was a better harness: a file that tracked every required feature and whether it had actually been tested, a running log of what had been done in each session, and a strict rule that each new session had to read that log before doing anything else. It is a concrete example of the broader harness engineering approach: changing the surrounding system so that known failure modes are less likely to recur.

Why this is not an abstract debate

There is a hard, measurable reason this shift matters right now, and it has nothing to do with hype. In the summer of 2025, researchers at MIT’s Project NANDA published a study called The GenAI Divide: State of AI in Business 2025, examining hundreds of publicly reported enterprise AI deployments alongside interviews and surveys of business leaders.⁹ Its most widely cited finding was that, despite an estimated thirty to forty billion dollars in enterprise investment in generative AI, roughly 95 per cent of the organisations studied had not yet demonstrated measurable profit-and-loss impact from their generative AI initiatives.

The report pointed toward problems of organisational integration, workflow and implementation rather than treating model capability alone as the explanation for failure; the companies that succeeded were not always using a more advanced model than the companies that failed. That distinction matters because harness engineering addresses precisely this surrounding layer: the workflows, tools, evaluations and feedback mechanisms that determine whether an AI system can function reliably in practice. Harness engineering is, in effect, the industry’s direct response to that gap. It is the name for the work required to move an AI system out of an impressive demonstration and into something an organisation can actually depend on.

How the world is responding, and who is pushing back

The response from major AI labs and tooling companies has been fast, and notably concrete. LangChain, the widely used framework for building AI agents, published a case study on February 17, 2026, reporting that its coding agent moved from 52.8 to 66.5 on Terminal Bench 2.0, moving from the Top 30 to the Top 5 at the time of publication, after changing the harness while keeping the underlying GPT-5.2-Codex model fixed. Every point of that improvement came from changing the harness around the model: how failures were traced, how the agent verified its own work, and how much context it was given about the surrounding codebase before it started.¹⁰

Datadog, the monitoring and observability company, published its own account in March 2026 of what it calls harness-first engineering, arguing that the fastest way to trust AI-generated code is not to read every line a human would have read, but to build automated checks precise enough to say, in seconds, whether the code is correct.¹¹ Other software engineering and governance organisations have also begun discussing harness quality alongside more familiar concerns such as testing, maintainability, security and observability.¹²

It is worth being honest that not everyone in the field agrees this is a genuinely new discipline rather than an old one wearing a new name. Some engineers and software practitioners have argued that harness engineering is less a wholly new discipline than a new label for familiar practices from software engineering, platform engineering and developer tooling.¹³

Birgitta Böckeler, a distinguished engineer at the consultancy Thoughtworks, published a more measured analysis in a pair of articles for Martin Fowler’s technology commentary series in February and April 2026, framing the elements of a harness as “guides” and “sensors” and arguing that a harness externalises some of the experience human developers normally bring to software development, while noting that its exact boundaries remain fluid and are still being worked out in practice.⁷ Both things can reasonably be true at once. The underlying engineering habits are not new. What is new, and what deserves attention, is that these habits are now being applied specifically and urgently to autonomous AI agents, at a moment when the gap between an impressive AI demonstration and a system an organisation can actually trust has become one of the defining problems in the industry.

Where India fits into this shift

Before turning to Jammu and Kashmir specifically, it is worth zooming out to India’s own position in this race, because the opportunity and the constraint both start at the national level.

India has built one of the world’s largest startup ecosystems. A written parliamentary reply put the number of startups formally recognised by the Department for Promotion of Industry and Internal Trade at 2,12,283 as of January 31, 2026, and government and industry trackers place the total above 2.4 lakh by the middle of 2026.¹⁴ The government’s flagship programme for artificial intelligence specifically, the IndiaAI Mission, was approved on March 7, 2024 with a total outlay of ₹10,371.92 crore, close to 1.25 billion US dollars, over five years.

Its seven pillars cover compute capacity, the IndiaAI Innovation Centre, datasets, application development, FutureSkills, startup financing, and safe and trusted AI.¹⁵ Investor confidence in this direction is visible in the funding numbers as well: domestic AI startup funding rose more than four times year on year in the first half of 2026, to $676 million, according to Inc42’s tracking, though analysts covering the same data have also asked, fairly, whether that pace is enough to let Indian AI companies compete globally rather than simply catch up to it.¹⁶

None of this national momentum guarantees that a discipline as specific as harness engineering will be adopted quickly or evenly across the country. The same national numbers that describe India’s opportunity also describe a real and specific skills challenge. Roughly 1.5 million engineers graduate from Indian institutions every year, but the harder problem is not simply that number; it is the gap between the size of that talent pool and the specialised skills that rapidly changing AI roles require.

Government estimates put India’s AI talent pool at roughly 600,000 to 650,000 in 2024, with demand projected to exceed 1.25 million AI professionals by 2027.¹⁷ That challenge is also unevenly distributed. India’s major technology hubs continue to concentrate much of the country’s advanced AI employment, infrastructure and experienced talent, while smaller cities are increasingly producing and attracting AI learners and entrepreneurs but often have less access to experienced practitioners working on production-scale systems. At the same time, more than half of all DPIIT-recognised startups nationally now come from Tier 2 and Tier 3 cities, official data shows, evidence that AI learning and entrepreneurship are spreading beyond India’s traditional technology hubs even as advanced, production-grade AI expertise remains concentrated in the largest ones.¹⁸

It is against this national backdrop, a country building serious public infrastructure and attracting real investment for artificial intelligence, while still working to close a specific and well-documented skills gap outside its major metros, that Jammu and Kashmir’s own position becomes easier to read clearly.

Why Jammu and Kashmir has a real opening

This is the part of the story that should matter most to a reader in Srinagar, Jammu, or the smaller towns of the Union Territory thinking about where to place their next few years of effort. Most previous shifts in the AI industry rewarded a scale that very few organisations outside a handful of countries could match: enormous compute budgets, access to proprietary training data, or teams of research scientists from a small number of universities. Harness engineering does not work that way. It is, at its core, a discipline of disciplined software practice: writing clear rules, building tests, logging failures, and refusing to let a known mistake happen twice. This is exactly the engineering culture that client services software teams build up over years of delivery work, regardless of where in the world they are based.

Jammu and Kashmir’s own technology and startup ecosystem has developed a considerably broader institutional base than is sometimes recognised, and has grown faster over the past several years than most readers outside the region realise. Government data placed before Parliament showed 1,628 startups recognised by DPIIT in the Union Territory as of June 30, 2026, up from 1,446 as of January 31, 2026, supported through central schemes such as the Fund of Funds for Startups and the Startup India Seed Fund Scheme. Srinagar accounted for the largest share, with 450 recognised startups, followed by Jammu with 431 and Baramulla with 130.¹⁹

The region’s incubation infrastructure includes facilities at the Islamic University of Science and Technology (IUST), SKUAST Kashmir and Jammu, the University of Kashmir, NIT Srinagar, IIT Jammu, Shri Mata Vaishno Devi University, the University of Jammu, the Jammu and Kashmir Entrepreneurship Development Institute (JKEDI), MIET and government polytechnic colleges in Jammu and Baramulla.²⁰ The SKIIE Centre at SKUAST-K in Srinagar was named Incubator of the Year at the Jammu and Kashmir Startup Mela in Jammu in January 2026.²¹

IUST’s Centre for Innovation and Entrepreneurship Development has been an active part of this ecosystem, and its Innovation Technology Business Incubator, associated with the centre and sponsored by the Department of Science and Technology, provides an additional avenue for supporting technology-based ventures.²² IIM Jammu’s Foundation for Entrepreneurship, Innovation and Skill Development operates business incubators in both Jammu and Srinagar.²³

The Jammu and Kashmir Startup Policy 2024-27 has set a formal target of supporting 2,000 startups over its three-year span.²⁴ Last month, in July 2026, JKEDI hosted the region’s largest-ever startup summit, ASCEND J&K 2026, at the Sher-i-Kashmir International Convention Centre in Srinagar, bringing together national investors, accelerators, and mentors. Twenty startups pitched to fourteen investors and investment networks; seven were shortlisted for further due diligence, and two secured concrete outcomes: Srinagar-based waste management startup EcoBuddy received an investment commitment in the range of ₹75 lakh to ₹1 crore from LetsVenture, and electric mobility startup Curve Electric secured funding support from Build3 Accelerator and angel investor Komal Nailwal.²⁵

None of this promises that harness engineering will be easy for local teams to adopt. It is evidence that the region already has an active, growing base of technical entrepreneurship this discipline could plug directly into, because the practice it demands, careful engineering rather than large-scale AI research, is one that experienced local software teams, including full-stack teams already delivering production platforms across health tech, education, and publishing, are well positioned to build on.

The constraints that deserve honesty, not optimism

It would be dishonest and unhelpful to any entrepreneur reading this piece to plan their next move, to stop at the opportunity without being equally clear about the real gap that stands in the way, because it is significant and it is documented, both nationally and locally.

As already noted, India’s AI talent shortfall is national in scope, and it lands hardest in Tier 2 and Tier 3 geographies, a category that includes most of Jammu and Kashmir’s towns and cities outside a small number of larger hubs.¹⁷ For engineers outside established AI communities, another challenge is simply finding and navigating the rapidly changing technical literature, documentation and practical examples surrounding agentic AI.

The region’s own incubation infrastructure, to its credit, is real and growing rather than notional, as the institutions named above make clear. They were built, sensibly, around the sectors that have historically driven entrepreneurship in the region: agriculture, general manufacturing, tourism-adjacent services, and broad-based digital ventures. Despite this growing institutional infrastructure, I have not found publicly documented evidence of a dedicated incubation or mentorship programme in J&K focused specifically on production-grade agentic AI systems or harness engineering. This is not a criticism of the institutions. It is an honest description of a gap. For a founder in Srinagar working on an AI agent for telemedicine, academic research or another specialised domain, publicly documented local mentorship focused specifically on production-grade agent harnesses appears limited. This mentorship gap is not unique to Jammu and Kashmir. It shows up consistently across India’s Tier 2 and Tier 3 startup ecosystems more broadly, where founders report that the hardest resource to find locally is not capital or ambition but experienced operators who have already solved the specific problem they are facing.²⁶

A realistic path through the gap, not around it

The honest response to a mentorship gap this specific is not to wait for it to close on its own, and it is also not to pretend the gap does not exist. A few developments already underway offer a genuinely usable path forward, and they are worth naming precisely because they are real programmes rather than aspirations.

The IndiaAI Mission’s FutureSkills pillar includes plans for Data and AI Labs in Tier 2 and Tier 3 cities, intended to give students and early founders in exactly these geographies hands-on exposure to real tools, real datasets, and structured problem-solving environments rather than theory alone.²⁷ Independent of government programmes, a pattern has also emerged across India’s smaller startup hubs of remote mentorship matching, connecting founders in cities such as Patna, Jodhpur, or Bhubaneswar directly with experienced operators and investors based in Delhi, Bengaluru, or Mumbai, without requiring the founder to relocate.²⁸ There is no structural reason this same model could not connect a Srinagar- or Jammu-based founder building an AI agent with an experienced harness engineer sitting anywhere else in the country or the world, given that so much of this work happens over video calls and shared code repositories rather than in person.

Perhaps the single most usable resource available to a resource-constrained team right now is one that costs nothing and requires no local expert at all: a particularly useful feature of this emerging field is that several companies at the frontier have publicly documented aspects of how they are building and evaluating agent systems. Anthropic’s account of how it solved the problem of AI agents losing track of long software projects, OpenAI’s account of building a million-line codebase almost entirely with agents, and LangChain’s own account of moving a coding agent from rank thirty to rank five on a public benchmark are all available to read in full, for free, by anyone with an internet connection. This is not a substitute for a mentor who has lived through the same mistakes locally, and it will not replace the value of a founder sitting across a table from someone who has done this work before. But it is a genuinely rare situation in a competitive technology industry, where the companies at the frontier are, for now, choosing to teach the discipline in public rather than guard it as a trade secret. For a founder in Jammu and Kashmir without access to a local harness engineering mentor, that open documentation is not a consolation prize. It is close to the same starting material an engineer at a well-funded firm in Bengaluru or San Francisco is reading this year.

Closing

None of this amounts to a promise that India, or Jammu and Kashmir specifically, will lead in agentic AI simply because a useful new discipline has arrived at a moment when both the country’s and the region’s own startup base happen to be growing. The gaps described here, in mentorship, in specialised AI talent, and in dedicated local expertise, are real, documented, and will not close on their own, whether one is looking at the national picture or the local one. Harness engineering may offer a different kind of opportunity from the compute-intensive competition that has dominated much of the AI industry: its emphasis is on disciplined software engineering, evaluation, tooling, observability and system design, rather than on training frontier foundation models from scratch. That is a fairer contest than the one the AI industry has offered smaller ecosystems for most of the past decade, and it is one that India’s growing base of technical entrepreneurs, and Jammu and Kashmir’s in particular, are genuinely positioned to compete in, provided the current mentorship gap is treated as a problem to be solved deliberately rather than waited out.

(The author is a scientist at the Salk Institute for Biological Studies, where his research focuses on neurodegeneration, with an emphasis on Alzheimer’s disease. Artificial intelligence is his personal interest rather than his field of active research. The views and opinions expressed in this article are his own and do not reflect the position of the Salk Institute or any other organisation with which he is affiliated.)

References

1. Mitchell Hashimoto, “My AI Adoption Journey,” mitchellh.com, February 5, 2026, https://mitchellh.com/writing/my-ai-adoption-journey.
2. Ryan Lopopolo, “Harness Engineering: Leveraging Codex in an Agent-First World,” OpenAI, February 11, 2026, https://openai.com/index/harness-engineering/.
3. Tobi Lütke, post on X, June 19, 2025, https://x.com/tobi/status/1935533422589399127.
4. Andrej Karpathy, post on X, June 25, 2025, https://x.com/karpathy/status/1937902205765607626.
5. Anthropic, “Effective Context Engineering for AI Agents,” Anthropic Engineering, September 29, 2025, https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents.
6. K. Hong, A. Troynikov, and J. Huber, “Context Rot: How Increasing Input Tokens Impacts LLM Performance,” Chroma Technical Report, July 2025, https://research.trychroma.com/context-rot.
7. Birgitta Böckeler, “Harness Engineering” and “Harness Engineering for Coding Agent Users,” martinfowler.com, February 17 and April 2, 2026.
8. Anthropic, “Effective Harnesses for Long-Running Agents,” Anthropic Engineering, November 26, 2025, https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents.
9. MIT Project NANDA, “The GenAI Divide: State of AI in Business 2025,” MIT Media Lab, July 2025, as reported in Fortune and Forbes coverage, August to October 2025.
10. Vivek Trivedy, “Improving Deep Agents with Harness Engineering,” LangChain Blog, February 17, 2026, https://blog.langchain.com/improving-deep-agents-with-harness-engineering/.
11. Datadog, “Closing the Verification Loop: Observability-Driven Harnesses for Building with Agents,” Datadog Blog, March 2026, https://www.datadoghq.com/blog/ai/harness-first-agents/.
12. Software Improvement Group, “What Is Harness Engineering?” softwareimprovementgroup.com, April 24, 2026.
13. Haverin (Substack), “What Is Harness Engineering? Why the AI Industry’s Newest Buzzword Is an Old Idea,” May 8, 2026.
14. Press Information Bureau, Government of India, written reply in the Lok Sabha on DPIIT-recognised startups, March 17, 2026, https://www.pib.gov.in/PressReleasePage.aspx?PRID=2241313; trend to mid-2026 as reported in TICE News, “India Crosses 2.40 Lakh DPIIT-Recognised Startups, Nears 24 Lakh Jobs Milestone,” July 1, 2026.
15. Press Information Bureau, Government of India, “Cabinet Approves Ambitious IndiaAI Mission to Strengthen the AI Innovation Ecosystem,” March 7, 2024, https://www.pib.gov.in/PressReleaseIframePage.aspx?PRID=2012355; outlay figure and pillar detail as reported in KNN India, “IndiaAI Mission to Drive AI Adoption Among MSMEs with Rs 10,371.92 Crore Outlay,” July 31, 2026.
16. Inc42, “Indian AI Startup Funding Soars Over 4X YoY in H1 2026, But Is It Enough to Compete Globally?” July 3, 2026.
17. NASSCOM–Deloitte India, “Advancing India’s AI Skills: Interventions and Programmes Needed,” 2024 estimate, as cited in Government of India (PIB) reporting, 2026.
18. Treelife, “India’s Revised Startup Recognition Framework 2026: What Every Founder Must Know,” April 10, 2026, citing DPIIT/Startup India data.
19. Department for Promotion of Industry and Internal Trade, written replies in the Rajya Sabha, as reported in Kashmir Life, “Jammu Kashmir Has 1400 Startup Ventures, Says Govt,” February 2026, and Daily Excelsior, “J&K gets Rs 950 cr industrial incentives in 5 yrs; 292 new units generate 29,000 jobs,” August 2026.
20. National Science and Technology Entrepreneurship Development Board, Department of Science and Technology, Government of India, directory of supported incubators in Jammu and Kashmir.
21. Kashmir Life, “Jammu Kashmir Has 1400 Startup Ventures, Says Govt,” February 2026.
22. Islamic University of Science and Technology, Centre for Innovation and Entrepreneurship Development, iust.ac.in; and Rising Kashmir, “Dr Singh to inaugurate two high-tech innovation centres at IUST,” August 22, 2025.
23. IIM Jammu, official website, iimj.ac.in.
24. JK Monitor, “Startup Ecosystem in Jammu & Kashmir: Technology, Innovation, and Growth Prospects,” December 29, 2025.
25. Kashmir Observer, “J&K’s Biggest Startup Summit Ends on High, Two Ventures Secure Funding,” July 24, 2026; and JK Monitor, “ASCEND J&K 2026 Concludes on a High Note; Two Startups Secure Funding, Investment Commitments,” July 2026.
26. Way2World, “Tier-2 Founder Challenges, Tier-3 Case Study, and VC Trends,” way2world.in, 2026.
27. IMPRI, “IndiaAI Mission’s FutureSkills Pillar: Building India’s AI-Ready Talent Pipeline,” impriindia.com, 2026.
28. Boot & Boost, “Tier-2 & Tier-3 Cities: The New Startup Goldmines in India,” bootboost.in.