TNS Agents Livestream: Jyoti Bansal, Harness

Key Concepts

• Continuous Delivery (CD): Automating the release of software to production.
• Continuous Integration (CI): Regularly merging code changes into a central repository.
• AI Agents: Autonomous software programs designed to perform specific tasks, often interconnected.
• Inner Loop/Outer Loop of Software Engineering: Inner loop refers to developers writing code; outer loop encompasses the entire delivery process from code commit to production.
• SDLC Knowledge Graph: A comprehensive, customer-specific data model of an organization's software development lifecycle, including infrastructure, services, environments, tools, and policies.
• Code Property Graph: A technology used to analyze code for exploitability and reachability of vulnerabilities.
• Multi-Cloud Platform (MCP): A flexible integration point for AI agents, acting as a "new API."
• Internal Developer Platform (IDP): A platform built for developers, often based on Backstage, to streamline workflows.
• Test Intelligence: An AI model that optimizes CI builds by identifying and running only necessary tests based on code changes.
• Consumption-Based Pricing: A business model where customers pay based on their usage of a service or product.
• Vertical AI Solutions: AI applications tailored for specific industries or domains.

---

Harness's Early Adoption and Vision for AI in Software Delivery

Yoti Bansal, CEO of Harness, highlights that Harness has been at the forefront of integrating Artificial Intelligence (AI) into software delivery since its inception. When Harness launched out of stealth in late 2017, its press release headline was "Harness brings artificial intelligence to continuous delivery." At that time, the AI used was primarily neural network-based machine learning, not Large Language Models (LLMs) or generative AI. An early application was using AI for Continuous Delivery (CD) to verify deployments and prevent breakage. By 2020, Harness introduced "Test Intelligence" for Continuous Integration (CI), an AI model that speeds up builds by 4x-5x by determining which tests need to run based on code changes, avoiding unnecessary test execution. Bansal emphasizes that AI is a means to solve problems, and not all problems require generative AI; sometimes, deterministic automation is faster and more appropriate.

The Challenge of the Outer Loop in Software Engineering

Bansal distinguishes between the "inner loop" (developers writing code) and the "outer loop" (the delivery process from code commit to production). He states that for engineering organizations at scale, 60-70% of time is spent in the outer loop, while only 30-40% is spent coding. This significant time investment in the outer loop is due to the complexity and sheer number of workflows involved, including:

• Testing: Integration, load, API, end-to-end, unit, resilience, chaos testing.
• Security: Code security, vulnerability scans, open-source library checks, API security, compliance approvals.
• Deployments: Code changes, feature flags, feature experiments, database changes, infrastructure as code, canary deployments, blue-green deployments, rollbacks, artifact management.
• Cost Optimization: Cloud cost management (FinOps). In total, there are 30-35 distinct workflows, making the outer loop a major bottleneck. Harness's goal is to streamline and automate these workflows, offering 16 different modules in its platform to address these challenges.

Evolution to Generative AI and Agent-Based Architecture

Harness's journey into generative AI began by simplifying configuration, such as creating pipelines or fixing security vulnerabilities using LLMs. Over time, it became clear that AI agents were the right solution for many problems. Harness has since built a library of interconnected AI agents, collectively known as "Harness AI." These agents operate at a top layer for broad DevOps, SRE, testing, FinOps, and security tasks, with purpose-built sub-agents handling smaller, specific tasks.

These agents leverage a "soft SDLC knowledge graph" created for each customer, which provides context about their infrastructure, service dependencies, environments, security tools, SLAs, and compliance standards. This contextual information allows agents to perform tasks accurately and efficiently, breaking down large problems into smaller, manageable tasks for specialized sub-agents.

Customer Experience and Trust in AI Agents

Harness's approach to customer interaction with AI agents is through a unified "Harness AI" interface, similar to how users interact with ChatGPT. Users don't directly see or manage individual agents; the system intelligently dispatches tasks to the appropriate agents. The user experience is personalized, with Harness AI suggesting likely tasks based on a user's past activities and project focus (e.g., deployment pipelines, cost optimization).

A critical aspect of customer adoption is trust. Harness addresses this by defining clear boundaries for AI agents:

• Pipeline Creation, Not Direct Production Deployment: Agents create auditable, reviewable, and deterministic production deployment pipelines, which humans then approve and manage. This ensures human-in-the-loop control and compliance.
• Troubleshooting Support: If a deployment fails, AI agents assist in troubleshooting, identifying configuration or code problems, and facilitating rollbacks. Bansal emphasizes that for production delivery, there is "no room for error," citing the CrowdStrike outage caused by a single line of bug. Unlike AI for coding (which can be 95-99% accurate), AI for software delivery requires near 100% accuracy due to the high impact of errors, necessitating deterministic processes and human oversight.

Prioritization of AI Agent Development

Harness prioritizes agent development based on the biggest pain points in software delivery:

1. Pipeline Setup: Creating complex, fully automated CI/CD pipelines for large enterprises, which must incorporate specific policies, security controls, and governance checks.
2. Testing: Helping engineers keep up with the increased code volume by providing AI agents for resilience, end-to-end, unit, and chaos testing.
3. Security: Addressing the "noise" of numerous vulnerabilities by prioritizing, automatically fixing, detecting, and determining the reachability and exploitability of vulnerabilities.

Strategic Acquisition: Qwak (formerly ShiftLeft)

Harness acquired Qwak (formerly ShiftLeft) to address the critical problem of managing and prioritizing security vulnerabilities. Qwak's "code property graph" technology helps identify which vulnerabilities are truly reachable and exploitable within a codebase, reducing the overwhelming noise from security scanners that report vulnerabilities in unused features of libraries. This technology is integrated into Harness AI to prioritize, detect, and automatically fix critical vulnerabilities.

Internal Transformation: Talent and Business Model

Harness, having a strong AI/ML foundation since 2017, had an advantage in transitioning to generative AI. Bansal notes that "almost all good engineers can become AI engineers" and are eager to learn. While strong data infrastructure and data scientists are crucial, Harness focuses on building deep domain expertise for software delivery problems rather than foundational AI models (they build on Anthropic, OpenAI, Gemini). The company's engineering culture has transformed to an "AI-first thinking."

From a business model perspective, Harness was already consumption-based, selling modular platform capabilities (e.g., CI, CD, FinOps, security testing) rather than seat-based licenses. AI is now natively built into the consumption pricing of these modules, not charged separately. This approach encourages engineers and product managers to design AI-native experiences without creating separate AI and non-AI versions.

Surprising Customer Adoption and Industry Trends

A key surprise for Harness has been the rapid and high adoption of AI integration, even in large, traditionally slow-moving enterprises, through their Multi-Cloud Platform (MCP). Bansal describes MCP as the "new API," offering flexibility for customers to integrate Harness AI into their internal toolchains and build their own autonomous agents. For example, a large airline customer uses Harness AI to create an internal AI toolchain for all its engineers.

The industry is realizing that while "AI for coding" increases code production, it doesn't automatically lead to faster shipping. The next challenge is "AI for software delivery" to manage testing, deployment, security, and compliance. Harness also offers capabilities for customers to create their own autonomous agents, such as one to upgrade a codebase library version, which can touch and modify customer code.

Bansal observes that AI is augmenting human labor rather than replacing it, leading to higher velocity and increased competitive pressure. The "appetite for doing more things" absorbs productivity gains, resulting in more work being done rather than fewer people.

Investment Landscape and the AI Bubble

As a co-founder of Unusual Ventures, Bansal sees 100% of investment pitches now incorporating AI. He is particularly interested in AI infrastructure and "vertical AI solutions" tailored for specific domains like healthcare, finance, or manufacturing.

Regarding the "AI bubble," Bansal acknowledges its existence but views it as potentially "not necessarily a bad thing." He compares it to the internet bubble: while expectations for rapid transformation (1 to 100 in two years) might be overblown, the underlying disruptive technology (1 to 100 in five years) is real. Many startups will fail, but significant winners will emerge, and incumbents will also adapt, leading to a transformative impact on the industry.

---

Synthesis/Conclusion

Harness, under Yoti Bansal's leadership, has consistently integrated AI into its software delivery platform, evolving from neural network-based machine learning to a sophisticated, agent-based generative AI architecture. The company's core mission is to automate and streamline the "outer loop" of software engineering, which consumes the majority of development time due to complex testing, security, deployment, and cost optimization workflows. Harness's AI agents, powered by customer-specific SDLC knowledge graphs, aim to simplify these processes while maintaining human oversight, auditability, and compliance—critical for production environments where even a single bug can have catastrophic consequences. Internally, Harness fosters an AI-first culture, enabling engineers to become AI practitioners and integrating AI natively into its consumption-based business model. The rapid adoption of Harness AI, even in large enterprises, underscores the industry's urgent need to leverage AI not just for code generation but for the entire software delivery pipeline, driving increased velocity and competitive pressure across the tech landscape. While acknowledging a "bubble" in AI investments, Bansal remains optimistic about AI's long-term disruptive potential, likening it to the internet's transformative impact.