Agents Prompt Humans in Reverse: A New Paradigm for Long-Horizon Financial Work

As AI agents evolve from "executing single-step instructions" to "advancing complex tasks over multiple days," the way financial work gets done is being redefined. Arca's AI lead recently shared a striking observation: in long-horizon financial work, agents are beginning to prompt humans in reverse—no longer is the human directing the AI step by step; instead, the AI proactively plans, advances, and raises the next items that need human confirmation. The core of this shift rests on persistent memory and a customized execution framework.

From "Tool Invocation" to "Continuous Presence"

Traditional AI assistants operate in a conversational mode: the human makes a request, the model responds, the task ends, and the context is wiped. This works well for short tasks, but much of financial work is inherently long-horizon—due diligence, portfolio monitoring, compliance review, earnings tracking—often spanning days or even weeks.

According to Arca's AI lead, the key bottleneck in long-horizon financial work isn't single-shot reasoning capability but memory continuity. The agent needs to remember where yesterday's analysis left off, which assumptions have been validated, and which data sources have shown anomalies. Persistent memory lets an agent work like a colleague who is continuously present, rather than starting from scratch each conversation.

This is also the precondition for "agents prompting humans in reverse": only with cross-session memory can AI assess current progress and proactively raise the next question.

Persistent Memory in the agent context refers to mechanisms by which a model saves and retrieves information beyond a single session, typically implemented through an external storage layer rather than relying on the model's own context window. Common approaches include: structured writes of task state, validated assumptions, and data-source anomaly records to a database or vector store, retrieved and injected on demand at the start of subsequent sessions. The counterpart is "in-session memory"—temporary information that exists only in the current context window and is lost when the session ends. The key design challenges of persistent memory are what to write, when to retrieve, and how to prevent stale or conflicting information from contaminating reasoning. For long-horizon financial tasks, memory accuracy directly determines whether the agent can continue advancing based on correct premises days later, so accompanying version management and a manual correction entry point are often required.

Custom Harness: Why General Frameworks Fall Short

The lead particularly emphasized the value of a customized harness. A harness can be understood as the execution scaffolding wrapped around the model, responsible for tool invocation, context management, state tracking, and error recovery.

General frameworks perform reasonably well in demo scenarios, but financial work has its particularities: diverse data source formats, strict compliance requirements, and intermediate conclusions that must be traceable. The value of a custom harness shows up in several ways—encoding domain rules into the execution flow to reduce the model's degrees of freedom on critical judgments; leaving an audit trail for every operation; and providing structured retry and validation mechanisms when model outputs are unstable.

In other words, model capability sets the ceiling, and the harness determines how much of that ceiling can actually be realized in real business processes.

Harness is a term borrowed from the testing and evaluation domain; in agent engineering it specifically refers to the orchestration layer wrapped around a base model. It doesn't change model weights but uses code logic to control what the model can do, which tools it can access, and how outputs are validated. A typical harness includes: tool registration and invocation routing, context assembly and trimming, loop control for multi-step tasks, failure retry and fallback strategies, and logging and audit records. General harnesses (such as common open-source agent frameworks) provide a generic loop, whereas a custom harness hard-codes domain constraints—for example, "financial data must come from designated data sources" and "key conclusions must include citations"—into the flow, thereby reducing the risk of the model improvising at critical junctures. This also explains why the same model may perform very differently in real business settings under different harnesses.

The Real Shape of Agent Workflows

The workflows mentioned here span units of "days." Such workflows differ fundamentally from common automation scripts:

• Phased progression: the agent breaks a long task into multiple stages, assessing at the end of each stage whether human intervention is needed
• Proactive questioning: when encountering information gaps or judgment disagreements, the agent doesn't silently assume but raises specific questions to humans
• Externalized state: progress, assumptions, and intermediate artifacts are persisted, so humans can inspect and correct them at any time

This brings a change in role division. Humans shift from "operators" to "reviewers and decision-makers," with their focus moving to setting goals, reviewing key nodes, and handling anomalies that agents proactively report.

Human-in-the-loop refers to a design pattern that presets human intervention points in automated flows. In long-horizon agent workflows, it typically isn't "ask a human at every step" but rather the agent proactively initiating inquiries under specific trigger conditions: an information gap that can't be filled from existing data sources, multiple candidate judgments with close confidence levels, or operations involving irreversible financial and compliance actions. Compared with traditional approval flows, the distinctive feature of human-in-the-loop is that the timing of intervention is dynamically judged by the agent rather than at fixed checkpoints, which requires the agent to have a reliable estimate of its own uncertainty. Poor design leads to two failure modes: too-frequent questioning makes humans the bottleneck, while too-infrequent questioning lets errors surface only days later. Therefore, the questioning threshold itself often needs to be an adjustable harness parameter, configured according to business risk levels.

Trade-offs and Open Questions in Practice

This model is not without costs. Persistent memory means more state management complexity, and a custom harness means higher development and maintenance investment. For scenarios with short task cycles and standardized processes, these investments may not pay off.

Another open question is the trust boundary. Can humans realistically review every intermediate decision an agent makes in a long-horizon task? If review density is too low, errors may surface only days later; if too high, the value of long-horizon automation is diminished. This requires finding a balance in process design rather than simply relying on model capability improvements.

The financial industry's requirement for traceability also makes audit logs and decision records mandatory components of the harness rather than optional ones.

Implications for Practitioners

Arca's practice points in a clear direction: the competitiveness of long-horizon agents increasingly depends on the engineering quality of memory architecture and execution frameworks, not just the strength of the underlying model. For financial teams evaluating agent deployment, three questions are worth prioritizing—which workflows have cycles long enough to justify investing in persistent memory infrastructure; which judgments must retain human approval nodes; and whether the existing process recording system can support the traceability of agent operations.

Agents prompting humans to work sounds like a role reversal, but in essence it's a redesign of the collaboration interface.

One-line Summary

Long-horizon financial agents' competitiveness depends on persistent memory and custom harnesses, with humans shifting to reviewers.

Paragraph Summary

Based on Arca AI lead's practical observations, the article discusses the shifting role of AI agents in long-horizon financial work: from passively executing single-step instructions to advancing tasks across days via persistent memory and proactively questioning humans. The author notes that the real bottleneck in long-horizon work isn't single-shot reasoning but memory continuity; general frameworks can't support financial scenarios, requiring custom harnesses that encode domain rules, audit trails, and error recovery into the execution flow. Agent workflows take the shape of phased progression, proactive questioning, and externalized state, shifting humans from operators to reviewers and decision-makers. The article also notes costs and open questions: rising state management and maintenance costs, questionable ROI for short-cycle standardized scenarios, and the need to balance review density between error exposure speed and automation value.