This article was co-authored by Rahul Vir and Reya Vir.
for the Token to work properly
We have officially passed the AI ​​prototyping phase. Building on concepts from Escaping the Prototype Mirage [1]product and engineering teams across the industry are now deploying applications that solve workflows previously handled by manual grinding. Building these independent agent models is now a breeze. It's as simple as using key concepts like recursive Agent Loops (Observe-Think-Act) to execute, setting up headless gateways to connect agents through chat applications, and relying on stored state that persists across restarts (as described in [1]). But their graduation from reliable products is another matter. The new frontier doesn't prove that agents can work, it proves that they can work profitably.
At the same time, internal metrics in companies such as “token maxing” (unforced token usage to obtain the best results) that were suitable for the prototyping phase are changing to measuring the ratio “value-to-token-sset” as a scale of agent products. After all, many products need to make profits and increase margins as they move from using cheap traditional computing (TradCompute) to solving user problems to using the same AI intelligence.
But models require freedom of thought and recent research has shown that the workflow of the testing agent goes beyond established methods, opens new methods, creates MCP tools, and builds infrastructure to solve the problem successfully in many cases. This brings up the question of balancing the need for an agency model with the economic reality of the cost of thinking.
Why Bonded Agents Fail to Meet
Agent harnesses store your work context and objectives in markup files (*.md), which usually do not represent a strict workflow, but rather express the goal or objective you want to achieve.
The Paradox of Mission Failure: In a study of agents solving complex problems, researchers found that giving strict guidelines, with too much delay when each action of the agent brings it closer to the goal, leads to local optimization and failure to achieve the goal. An example from Professor Jeff Clune's research on open-ended agent learning illustrates this well: an agent in a maze, when constantly rewarded for looking for a direct exit, will repeatedly crash into walls and get stuck in an ideal location, never reaching the end. [2].
Unrestricted Harness Power: Contemporary agent harnesses such as Google's Antigravity and Anthropic's Claude Code have worked very well because they allow agents to create, plan, perform complex tasks, and build their tools without strict human management. They succeed because they are given the freedom to explore roundabout ways.
Consider an edge case in a typical medical care workflow: if we force a healthcare agent strictly to follow a pre-defined schedule flow, it happens in the real world. If a patient complains of chest pain during that routine meal, the agent's Agentic Loop should have the autonomy to recognize the urgency, abandon the schedule flow, and initiate a safety escalation. It should use what we previously described as the “No Response Signal'' to suppress the booking conversation and deliver the context directly to a human nurse. [1]. The most robust prototypes fail this test spectacularly because they cannot adapt to critical, out-of-bounds conditions.
Endless Goal Searching is Expensive
While provisioning is important for a solution initially, implementing a full-fledged search for every request in the workflow can lead to large and unsustainable token consumption. At this stage the agent has found a valid path and this path naturally allows it to re-evaluate or “visualize” the workflow structure. While this may be self-correcting, doing the same for the same request destroys the token economy of the business.
For example, medical intake workflows and critical situations that need to be escalated can be studied over time. Clinical or solution provider workflows will end up in deterministic methods for the most part, leaving some autonomy to rare outliers and complex situations.
Architectural Solutions with Early Commitment and Deterministic Replay
Early Commitment has shown promise in solving systematic problems and can also be used in agent workflows. [3]. It involves first classifying the problem, say by programming the system information to require the model to generate a specific classification marker. By forcing the agent to isolate the nature of the problem and establish constraints before generating an actionable idea, you prevent the agent from seeing things that are missing or exploring endless paths. This cuts through the noise and focuses the agent only on execution instead of continuous evaluation.
For example, in a telehealth triage workflow, we can enforce Early Commitment by requiring the agent to specifically classify the encounter as “prescription refill” before taking any action. Once committed to this specific threshold, the agent limits its tool calls strictly to the pharmacy database, completely bypassing the expensive, open-ended diagnostic methods that may be wandering around trying to diagnose a patient.
A recent study by Wang, X., et al. introduces the LOOP Skill Engine Framework, which takes early commitment at the infrastructure level by using a single recording and deterministic replay paradigm [4]. The agent can automatically check once using full logic, and the system combines that successful lead into a branchless recipe. In all future runs, LLM can be bypassed, guaranteeing artificial determinism and cutting off token usage in more than 93.3% of daily transactions, and up to 99.98% in high-frequency usage. This concept can be extended to agent workflow.
Consider the production of daily clinical compliance reports or regular post-discharge summaries, which are stable, repetitive tasks. From assessment and rapid completion to a decision framework, the agent must consider extracting complex data from the Electronic Health Record simultaneously. For the next hundred patients discharged through the same procedure, the system uses that branch-free recipe, reliably changing the patient's priorities and dates without asking the LLM. This ensures zero forgotten data in repetitive healthcare operations while increasing token efficiency.
ML practitioners need to make a call between pure deterministic replay (like LOOP) that maximizes token savings, and a hybrid approach (keeping the checked path in the SKILL.md file). A hybrid approach trades some of that token savings for a more efficient targeted exchange, yet leaves enough flexibility to adapt itself to a changing infrastructure. Whether this skill file is updated manually or through a self-developing machine, maintaining this mindset ensures long-term adaptability and resilience. For example, if the database structure changes, the agent is able to update the SQL queries and retrieve the information.
Conclusion: An Explore-Commit-Measure ML Pipeline
ML developers and product managers must adapt their applications to take advantage of the greater intelligence of autonomous agents and adopt agentless harnesses for early problem detection and complex, one-off scenarios. This presents the best solutions without using an expensive reinforcement learning cycle (which is often prevented by lack of technology, platform constraints, training costs or closed models).
Once we've found a way around it, the token economy of structured and repeatable tasks requires us to force early commits to fast builds, using a deterministic replay architecture to maintain a cache of execution methods.
As a measure of agent productivity, we must move performance metrics away from simple measures of success, instead toward token interactions and value-per-token generated.
References
- Vir, R., & Vir, R. (2026, March 4). Escaping the mirage of the prototype: Why business AI stalls. About Data Science.
- Clune, J. (2025, February 12). Guest lecture 6 CS329A by Prof. Jeff Clune: Open Agent Learning in the Age of Basic Models [Video]. YouTube.
- Vir, R. (2026, January 1). Why early commitment helps AI solve systematic problems. Inside AI.
- Wang, X., Yu, K., Liang, X., Wang, L., & Han, C. (2026). Good to go: LOOP skill engine that achieves 99% success and reduces token consumption by 99% with single recording and decisive replay. arXiv.
