# Beyond the Harness: A Journey Towards Adaptive Engineering Rajiv Chandegra, Annicha Labs | AI Engineer World's Fair 2026 | 37:01 Source: https://www.youtube.com/watch?v=qdZzND79mcg Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/beyond-the-harness-a-journey-towards-adaptive-engineering Published: 2026-07-07 Tags: agents, guardrails, multi-agent, workflows ## TL;DR - Fixed harnesses make AI engineering reliable and auditable, but they suppress the variation that produces novelty. - Real-world problems are complex systems of changing relationships, so they require agents and structures that can adapt during execution. - Adaptive engineering gives engineers control over constraints while allowing the multi-agent harness to emerge, stabilize, and reorganize in response to its environment. ## Summary Rajiv Chandegra argues that current AI engineering depends on fixed harnesses. These harnesses define roles, tools, prompts, sequencing, memory, and handoffs before a run starts. That approach works well for predictable, complicated problems, but it becomes brittle when AI interacts with people, institutions, and the physical world. Chandegra uses complexity science to distinguish complicated systems, such as clocks and aircraft, from complex systems, such as markets, organizations, and bird flocks. In complex systems, local interactions create patterns that cannot be designed from the parts alone. Adaptive engineering responds by defining constraints rather than prescribing the full structure. Agents interact, specialize, form clusters, create conventions, and reorganize as the environment changes. The engineer senses and responds to the emerging harness instead of repeatedly rebuilding it. Chandegra is also clear about the risks: drift, monoculture, reduced legibility, and less predictability before runtime. He presents adaptive engineering as a complement to fixed engineering, with different use cases. ## Key ideas ### Fixed harnesses make stateless models useful through predetermined structure [00:24](https://www.youtube.com/watch?v=qdZzND79mcg&t=24s) Current AI engineering uses a harness to steer an otherwise stateless large language model. The harness can include vendor system prompts, agents.md or Claude.md files loaded at the start of a session, tool calling, specialized agents, skills, rules, loops, and human review. Roles, outputs, sequencing protocols, and memory creation are decided before the engineering run begins. This makes the process reliable, auditable, and traceable. The same input tends to produce a similar set of outputs, changes can be inspected, and failures can be followed back to their source. ### Factory-style harnesses work for fixed problems and suppress novelty in moving ones [07:35](https://www.youtube.com/watch?v=qdZzND79mcg&t=455s) Chandegra compares fixed harness engineering with a factory assembly line. Each agent has one job, a fixed place, a sequence, and perhaps a defined handoff. This gives accuracy, speed, reproducibility, and certifiable behavior in closed, deterministic systems with well-defined products. The cost is that reliability comes from suppressing variance, while novelty needs variance. As models improve, a harness designed today may become unnecessary or outdated soon. When unexpected situations appear, humans patch the harness with more rules until its structure becomes more complicated than the problem. ### Complex real-world problems are changing relationships rather than collections of parts [10:40](https://www.youtube.com/watch?v=qdZzND79mcg&t=640s) The reductionist view treats the world as stable parts whose relationships are secondary. The systems view puts relationships first. A stable thing becomes a slow pattern in an ongoing process, like a flame that exists only while its process continues. Local interactions can create properties that no individual component has, such as water's wetness or a bird flock's coordinated movement. Russell Ackoff's term 'mess' describes a dynamic situation where problems interact and change. Such situations cannot be decomposed into tidy boxes because the relationships are part of the problem. ### Complex systems need probing and adaptation instead of prediction from isolated parts [16:08](https://www.youtube.com/watch?v=qdZzND79mcg&t=968s) Chandegra separates complicated problems from complex problems. A jumbo jet or clock has passive parts that experts can analyze, plan, document, and predict. A flock, market, or human organization has parts that continually interact and adapt, so the whole cannot be derived from the parts. The appropriate method is to probe, sense, and respond. Complex systems contain diverse agents, local interaction, recursive learning, and adaptation to the environment. Their interactions can produce emergence, while attractors create stable patterns at the system level even as local activity keeps changing. ### Adaptive engineering lets the harness emerge during the engineering run [20:59](https://www.youtube.com/watch?v=qdZzND79mcg&t=1259s) Adaptive engineering starts from the idea that a harness should adapt mid-flow when AI operates in a changing social or physical environment. Chandegra defines it as designing constraints so that the harness emerges, stabilizes, and adapts in ways that could not be specified in advance. The harness becomes the ongoing output rather than the input. Agents interact in response to a goal and environment, then form an organization that fits the current situation. In his analogy, agents create the harness through their relationships in the same way birds create a flock. ### Agent specialization and governance can arise from local interaction [23:01](https://www.youtube.com/watch?v=qdZzND79mcg&t=1381s) In Chandegra's model, initially similar agents interact without fixed differentiation. As coupling builds, a system begins to form. Environmental pressure rewards small differences between redundant agents, so those differences become specialized roles or capabilities. Agents then form clusters, and the system creates its own boundaries. Repeated interaction can produce shared norms, protocols, and governance without a central governor. The resulting order remains stable until the environment changes, at which point the system can restructure itself. The engineer's control shifts toward the rate and conditions of interaction. ### The engineer designs constraints and responds to the harness that emerges [26:12](https://www.youtube.com/watch?v=qdZzND79mcg&t=1572s) Adaptive engineering does not remove the engineer. It changes what the engineer decides. The engineer sets the rules of the game and gives agents room to explore rather than prescribing the desired result. Useful design questions include whether to give agents more freedom or more guardrails, whether to reward coherence toward a goal or penalize movement outside a boundary, and how quickly agents should couple. Once an emergent harness takes shape, the engineer senses and responds to it during the run instead of stopping and starting over. ### Fixed and adaptive engineering form a continuum with different failure modes [27:29](https://www.youtube.com/watch?v=qdZzND79mcg&t=1649s) Chandegra does not present adaptive engineering as universally better. Fixed engineering prescribes the structure and rarely changes it during execution. Adaptive engineering lets agents interact and allows the guiding structure to change during the run. He distinguishes adaptive design-time tools from systems that reorganize at runtime, citing Pi as an example of the former and saying it does not claim to be a self-organizing multi-agent system. Adaptive systems can settle into suboptimal attractors, drift without selection pressure, develop monoculture, lose legibility, and sacrifice predictability before runtime. ## Notable quotes - "The factory method is the right answer to a fixed problem and the wrong answer to a moving problem." (10:12) - "The flock lives in the relationships, not the parts." (13:56) - "The harness becomes the ongoing output rather than the input." (21:43) - "You don't abolish the engineer, you're just relocating the emphasis of engineering." (26:38) - "The limiting factor which is the case now, but probably more so in the future, is not going to be the strength of the model. It's going to be the adaptability of the harness." (35:41) ## Tools & references mentioned - Claude Code - Cursor - Codex - Pi - LangChain - Hermes AI - Klein - Goose - Russell Ackoff - Taylorism ## Who should watch - You are building agent workflows with fixed roles and handoffs, and want to understand where that structure may fail as the problem changes. - Your AI system will interact with people, institutions, or the physical world rather than staying inside a software-only environment. - You are exploring multi-agent coordination and need a clear account of its possible benefits, along with risks such as drift, monoculture, and weak legibility.