This text was co-authored by Reya Vir and Rahul Vir.
has essentially modified within the GenAI period. With the ubiquity of vibe coding instruments and agent-first IDEs like Google’s Antigravity, creating new functions has by no means been sooner. Additional, the highly effective ideas impressed by viral open-source frameworks like OpenClaw are enabling the creation of autonomous techniques. 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 through 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 a substitute of their normal chatty nature.
Constructing autonomous brokers has been a breeze. However the query stays: if constructing is so frictionless at this time, 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 functions 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 achieve funding, nevertheless it not often survives in the true world.
That is largely on account of a sudden improve in vibe coding that prioritizes speedy experimentation over rigorous engineering. These instruments are superb at creating demos, however with out structural self-discipline, the ensuing code lacks the potential 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 difficulty 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 now we have a Affected person Consumption Agent designed to triage sufferers, confirm insurance coverage, and schedule appointments. In a vibe-coded demo, it handles normal 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 setting is stateful and messy. If a affected person mentions chest ache halfway by a routine consumption, the agent’s Agentic Loop should immediately acknowledge the urgency, abandon the scheduling circulate, 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 take a look at spectacularly.
Right now, 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 fact of the manufacturing setting.
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 on account 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 embody:
- Unknown Reliability: Most brokers fall wanting the strict Service Stage Agreements (SLAs) enterprise use calls for. Because the errors inside single- or multi-agent techniques 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 cheap start line contemplating using brokers in these extremely specialised domains the place public benchmarks are inadequate, the method is neither scalable nor maintainable. Transferring to structured evals and LLM-as-a-Choose 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 research its personal failures logs and adapt. Its inflexible immediate chains break as quickly because the setting 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 setting and repeatedly adapt to deal with the challenges with out fixed human intervention–they have to be directionally conscious of the true optimization aim.
OKRs present a superior goal to realize (e.g., Cut back important affected person wait occasions by 20%) quite than an intermediate aim metric (e.g., Course of 50 consumption varieties an hour). By understanding the OKR, our Affected person Consumption Agent can thus proactively see indicators that run counter to the affected person wait time aim and handle them with minimal human involvement.
Latest analysis from Berkeley CMR frames this within the principal-agent idea. The “Principal” is the stakeholder answerable for the OKR. Success will depend on delegating authority to the agent in a approach 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 circumstances with strict guardrails (e.g., the agent solely handles routine physicals and primary insurance coverage verification).
- Escalation: The agent acknowledges edge circumstances (e.g., conflicting signs) and escalates to human triage nurses quite than guessing.
- Evolution: Because the agent good points higher information 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 important to remodel these prototypes into true merchandise that evolve over time. We’ve to grasp that agentic functions must be developed, developed, and maintained to develop from mere assistants to autonomous entities–identical to software program functions. Vibe-coded mirages will not be 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 carry product alignment and engineering self-discipline to the event of those brokers. We’ve to construct techniques to fight the particular methods these fashions wrestle, resembling these recognized in 9 important failure patterns.
Over the subsequent few weeks, this sequence will information you thru the technical pillars required to remodel your enterprise.
- Reliability: Transferring from “Vibes” to Golden Datasets and LLM-as-a-Choose (so our Affected person Consumption Agent could 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 through information lineage and circulate management.
- Efficiency: Attaining agent efficiency at scale to enhance productiveness.
The journey from a “Prototype” to “Deployed” isn’t about fixing bugs; it’s about constructing a essentially 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. Knowledge, 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 Overview. https://cmr.berkeley.edu/2025/07/rethinking-ai-agents-a-principal-agent-perspective/
- Vir, R., Columbia DAPLab. (2026, January 8). 9 Crucial Failure Patterns of Coding Brokers. Knowledge, Brokers, and Processes Lab, Columbia College. https://daplab.cs.columbia.edu/common/2026/01/08/9-critical-failure-patterns-of-coding-agents.html
All photos generated by Nano Banana 2
