Stop Building More AI Agents. Build Reliable Loops Instead

The latest advice in AI development sounds deliberately provocative: stop creating more agents and start writing loops. Boris Cherny, the creator and head of Claude Code, has described a shift from humans prompting agents to agents prompting other agents that continue working in the background. The idea deserves attention—but not because loops magically become self-improving employees. Their value comes from turning isolated AI actions into repeatable, measurable systems.

What is an AI agent loop?

An agent answers a request by reasoning, using tools, and taking actions. A loop adds a controlled cycle around that work. It defines an objective, plans the next step, executes it, inspects the resulting state, and decides whether to continue, revise, escalate, or stop.

A practical loop might repeatedly inspect a codebase for duplicated abstractions, propose one improvement, run tests, and open a pull request only when the change passes its checks. TechCrunch’s account of Cherny’s workflow describes similar agents continuously looking for architectural improvements and submitting changes for review.

This is less mysterious than the terminology suggests. The essential sequence is:

1. Define a bounded objective and the state the system may change.
2. Select the next action using an explicit plan or policy.
3. Execute through approved tools.
4. Verify the outcome against measurable criteria.
5. Record the result and decide whether to repeat or stop.

The loop is the operating structure around the model. It is not evidence that the model learns by itself. Unless the system deliberately writes validated information into memory, updates its instructions, or changes another controlled artifact, the next run does not automatically become better.

Reliable loops need more than a planner and a builder

The planner–builder–verifier pattern is a useful starting point. The planner breaks an objective into steps. The builder calls tools and changes the environment. The verifier checks whether the result meets the intended outcome. For long-running work, the system also needs durable state so that a later execution can understand what was attempted and what remains.

Anthropic’s work on long-running agent harnesses makes the limitation clear: a frontier model placed in a loop with only a broad prompt still struggles to deliver production-quality work across multiple context windows. Effective systems leave structured artifacts for the next session, make incremental progress, and preserve a clear account of the current state.

A production loop therefore needs several additional controls:

• a scope that limits which systems, records, and actions are available;
• success criteria that can be evaluated from the actual outcome;
• budgets for time, tokens, retries, and tool calls;
• stop conditions for completion, uncertainty, repeated failure, and rising cost;
• approval gates before external messages, payments, deletions, or other consequential actions;
• traceable records of inputs, decisions, tool calls, and changes.

Without those controls, “keep working until it is done” can become “keep spending while errors compound.”

Verification is the engine of improvement

The verifier is what separates a useful loop from repeated generation. It should evaluate the resulting state, not simply ask another model whether the first model’s answer looks convincing. Tests, database checks, policy constraints, source citations, latency, cost, and calibrated human review can all contribute evidence.

Anthropic’s guidance on evaluating AI agents recommends combining code-based, model-based, and human graders. It also distinguishes the transcript—what the agent said and did—from the outcome—what actually changed in the environment. That distinction matters: an agent can confidently report that a task succeeded when the expected record, file, reservation, or approval does not exist.

Evaluation also enables genuine improvement over time. Teams can collect representative failures, turn them into test cases, compare versions, and update instructions or tools when results regress. The loop does not autonomously learn in the machine-learning sense; the surrounding system creates a disciplined feedback process.

Not every task should become a loop

Continuous agents consume more tokens, create more opportunities for drift, and expand the surface area for permission mistakes. Anthropic’s earlier agent-design guidance recommends starting with the simplest solution and increasing complexity only when the expected value justifies the added cost and latency.

Loops work best when progress is observable, outcomes are testable, and another useful iteration is likely. Code maintenance, data-quality monitoring, reconciliation, research updates, and controlled content operations can fit that pattern. Ambiguous decisions, irreversible actions, and work requiring accountable expert judgment should retain meaningful human control.

Every loop is limited by its enterprise context

An agent loop can verify code with tests because the environment exposes clear evidence. Enterprise knowledge work is harder. A loop may retrieve an obsolete policy, mistake an opinion for a rule, lose a document’s jurisdiction, or act on data that the user was not permitted to access. Repeating the cycle does not repair missing context; it can repeat the same mistake more efficiently.

This is where the topic connects directly to Goldlayer. Reliable loops need governed, traceable, AI-ready data: structured sources, useful metadata, freshness, authority, permissions, and lineage. When enterprise documents disagree, the system needs enough context to clarify or constrain the task before another iteration begins.

Goldlayer’s direction is to prepare that knowledge layer inside company-controlled infrastructure so agents, loops, search, and analytics can share a more reliable foundation. The future may involve fewer manually written prompts and more systems that keep working. But the winning architecture will not be the loop that runs the longest. It will be the one that knows what evidence to trust, what it is allowed to change, how success is measured, and when to stop.