This text was co-authored by Reya Vir and Rahul Vir.
has basically modified within the GenAI period. With the ubiquity of vibe coding instruments and agent-first IDEs like Google’s Antigravity, growing new purposes has by no means been quicker. Additional, the highly effective ideas impressed by viral open-source frameworks like OpenClaw are enabling the creation of autonomous programs. We will drop brokers into safe Harnesses, present them with executable Python Abilities, and outline their System Personas in easy Markdown information. We use the recursive Agentic Loop (Observe-Assume-Act) for execution, arrange headless Gateways to attach them by way of chat apps, and depend on Molt State to persist reminiscence throughout reboots as brokers self-improve. We even give them a No-Reply Token to allow them to output silence as an alternative of their normal chatty nature.
Constructing autonomous brokers has been a breeze. However the query stays: if constructing is so frictionless in the present day, why are enterprises seeing a flood of prototypes and a remarkably small fraction of them graduating to precise merchandise?
1. The Phantasm of Success:
In my discussions with enterprise leaders, I see innumerable prototypes developed throughout groups, proving that there’s immense bottom-up curiosity in reworking drained, inflexible software program purposes into assistive and absolutely automated brokers. Nonetheless, this early success is misleading. An agent might carry out brilliantly in a Jupyter pocket book or a staged demo, producing sufficient pleasure to showcase engineering experience and acquire funding, but it surely hardly ever survives in the true world.
That is largely as a consequence of a sudden improve in vibe coding that prioritizes fast experimentation over rigorous engineering. These instruments are superb at growing demos, however with out structural self-discipline, the ensuing code lacks the aptitude and reliability to construct a production-grade product [Why Vibe Coding Fails]. As soon as the engineers return to their day jobs, the prototype is deserted and it begins to decay, identical to unmaintained software program.
The truth is, the maintainability concern runs deeper. Whereas people are completely able to adapting to the pure evolution of workflows, the brokers aren’t. A refined enterprise course of shift or an underlying mannequin change can render the agent unusable.
A Healthcare Instance: Let’s say we’ve got a Affected person Consumption Agent designed to triage sufferers, confirm insurance coverage, and schedule appointments. In a vibe-coded demo, it handles commonplace check-ups completely. Utilizing a Gateway, it chats with sufferers utilizing textual content messaging. It makes use of primary Abilities to entry the insurance coverage API, and its System Persona units a well mannered, scientific tone. However in a dwell clinic, the surroundings is stateful and messy. If a affected person mentions chest ache halfway by way of a routine consumption, the agent’s Agentic Loop should immediately acknowledge the urgency, abandon the scheduling circulation, and set off a security escalation. It ought to make the most of the No-Reply Token to suppress reserving chatter whereas routing the context to a human nurse. Most prototypes fail this check spectacularly.
Immediately, a overwhelming majority of promising initiatives are chasing a “Prototype Mirage”–an limitless stream of proof-of-concept brokers that seem productive in early trials however fade away after they face the truth of the manufacturing surroundings.
2. Defining The Prototype Mirage
The Prototype Mirage is a phenomenon the place enterprises measure success primarily based on the success of demos and early trials, solely to see them fail in manufacturing as a consequence of reliability points, excessive latency, unmanageable prices, and a basic lack of belief. Nonetheless, this isn’t a bug that may be patched, however a systemic failure of structure.
The important thing signs embrace:
- Unknown Reliability: Most brokers fall in need of the strict Service Stage Agreements (SLAs) enterprise use calls for. Because the errors inside single- or multi-agent programs compound with each motion (aka stochastic decay), builders restrict their company. Instance: If the Affected person Consumption Agent depends on a Shared State Ledger to coordinate between a “Scheduling Sub-Agent” and an “Insurance coverage Sub-Agent,” a hallucination at step 12 of a 15-step insurance coverage verification course of derails the entire workflow. A latest examine reveals that 68% of manufacturing brokers are intentionally restricted to 10 steps or fewer to forestall derailment.
- Analysis Brittleness: Reliability stays an unknown variable as a result of 74% of brokers depend on human-in-the-loop (HITL) analysis. Whereas it is a affordable place to begin contemplating the usage of brokers in these extremely specialised domains the place public benchmarks are inadequate, the strategy is neither scalable nor maintainable. Shifting to structured evals and LLM-as-a-Decide is the one sustainable path ahead (Pan et al., 2025).
- Context Drift: Brokers are sometimes constructed to snapshot legacy human workflows. Nonetheless, enterprise processes shift naturally. Instance: If the hospital updates its accepted Medicaid tiers, the agent lacks the Introspection or Metacognitive Loop to investigate its personal failures logs and adapt. Its inflexible immediate chains break as quickly because the surroundings diverges from the coaching context, rendering the agent out of date.
3. Alignment to Enterprise OKRs
Each enterprise operates on a set of outlined Aims and Key Outcomes (OKRs). To interrupt out of this phantasm, we should view these brokers as entities chartered to optimize for particular enterprise metrics.
As we goal for larger autonomy–permitting brokers to grasp the surroundings and repeatedly adapt to handle the challenges with out fixed human intervention–they have to be directionally conscious of the true optimization objective.
OKRs present a superior goal to attain (e.g., Cut back essential affected person wait occasions by 20%) slightly than an intermediate objective metric (e.g., Course of 50 consumption types an hour). By understanding the OKR, our Affected person Consumption Agent can thus proactively see alerts that run counter to the affected person wait time objective and handle them with minimal human involvement.
Current analysis from Berkeley CMR frames this within the principal-agent concept. The “Principal” is the stakeholder liable for the OKR. Success is dependent upon delegating authority to the agent in a manner that aligns incentives, making certain it acts within the Principal’s curiosity even when operating unobserved.
Nonetheless, autonomy is earned, not granted on day one. Success follows a Guided Autonomy mannequin:
- Identified Knowns: Begin with skilled use instances with strict guardrails (e.g., the agent solely handles routine physicals and primary insurance coverage verification).
- Escalation: The agent acknowledges edge instances (e.g., conflicting signs) and escalates to human triage nurses slightly than guessing.
- Evolution: Because the agent positive aspects higher knowledge lineage and demonstrates alignment with the OKRs, larger company is granted (e.g., dealing with specialist referrals).
4. Path Ahead
A cautious long-term technique is crucial to remodel these prototypes into true merchandise that evolve over time. Now we have to grasp that agentic purposes must be developed, advanced, and maintained to develop from mere assistants to autonomous entities–identical to software program purposes. Vibe-coded mirages are usually not merchandise, and also you shouldn’t belief anybody who says in any other case. They’re merely proof-of-concepts for early suggestions.
To flee this phantasm and obtain actual success, we should deliver product alignment and engineering self-discipline to the event of those brokers. Now we have to construct programs to fight the particular methods these fashions battle, reminiscent of these recognized in 9 essential failure patterns.
Over the subsequent few weeks, this collection will information you thru the technical pillars required to remodel your enterprise.
- Reliability: Shifting from “Vibes” to Golden Datasets and LLM-as-a-Decide (so our Affected person Consumption Agent might be repeatedly examined in opposition to 1000’s of simulated complicated affected person histories).
- Economics: Mastering Token Economics to optimize the price of agentic workflows.
- Security: Implementing Agentic Security by way of knowledge lineage and circulation management.
- Efficiency: Reaching agent efficiency at scale to enhance productiveness.
The journey from a “Prototype” to “Deployed” will not be about fixing bugs; it’s about constructing a basically higher structure.
References
- Vir, R., Ma J., Sahni R., Chilton L., Wu, E., Yu Z., Columbia DAPLab. (2026, January 7). Why Vibe Coding Fails and The best way to Repair It. Information, Brokers, and Processes Lab, Columbia College. https://daplab.cs.columbia.edu/common/2026/01/07/why-vibe-coding-fails-and-how-to-fix-it.html
- Pan, M. Z., Arabzadeh, N., Cogo, R., Zhu, Y., Xiong, A., Agrawal, L. A., … & Ellis, M. (2025). Measuring Brokers in Manufacturing. arXiv. https://arxiv.org/abs/2512.04123
- Jarrahi, M. H., & Ritala, P. (2025, July 23). Rethinking AI Brokers: A Principal-Agent Perspective. Berkeley California Administration Evaluation. https://cmr.berkeley.edu/2025/07/rethinking-ai-agents-a-principal-agent-perspective/
- Vir, R., Columbia DAPLab. (2026, January 8). 9 Essential Failure Patterns of Coding Brokers. Information, Brokers, and Processes Lab, Columbia College. https://daplab.cs.columbia.edu/common/2026/01/08/9-critical-failure-patterns-of-coding-agents.html
All pictures generated by Nano Banana 2
