Start with the work
An agent demonstration is easy to admire: a request goes in, a sequence of actions follows, and a finished result appears. The less visible question is what happens when the request is ambiguous, a tool returns the wrong information, or an action succeeds only halfway. A useful agent needs a design for those cases as well.
We use “agent” to mean a system in which a model can choose its next action in response to what it observes. A workflow can instead specify most of the sequence in advance. Anthropic’s foundational engineering essay makes a similar distinction and recommends increasing complexity only when the task warrants it. Its tooling examples have evolved since its original publication, so we draw on the architectural distinction rather than treating it as a current product guide. [1]
Our design recommendation begins with a description of the work, written without the word “agent.” What outcome does the person need? Which decisions vary from case to case? What evidence is available? The architecture should follow those answers.
Define an operating envelope
Consider a fictional assistant that prepares a supplier comparison. It may search an approved document collection, extract prices, and assemble a draft. That does not imply permission to contact suppliers, accept terms, or spend money. A goal describes the desired result; an operating envelope describes what the system may do to reach it.
Write that envelope in concrete terms. Identify the accessible data, the permitted tools, the maximum effort, and the actions that require review. Separate reading a record from editing it, and drafting a message from sending it. Enforce consequential limits in the surrounding software, rather than relying exclusively on instructions to the model.
The boundary should be understandable to the person using the system. A vague request to approve “continuing” is less useful than a preview of the exact proposed change. Review works better when the person can inspect what will happen and why the agent believes it is necessary.
Give feedback a job
A tool response should help the agent decide whether its last step worked. If a document was created, can the system inspect it? If a record was changed, can the new state be read back? If a search produced no useful evidence, is that absence distinguishable from a service error?
Anthropic’s essay emphasizes using environmental feedback during execution and setting stopping conditions. [1] Our interpretation is that feedback should be designed as part of the task contract. Merely allowing more attempts does not explain how the system will recognize progress.
For the supplier example, progress might mean filling a required comparison field with a source-backed value. A missing price should remain missing, with a reason, rather than being replaced by a plausible estimate. The finished report should make that uncertainty visible.
Plan for partial success
Real work often ends somewhere between completion and failure. An agent may gather all the sources but fail to format the report. It may update two records before a third rejects the request. The product needs to show what changed, what remains, and what can safely be retried.
A practical approach is to assign explicit states to consequential actions: proposed, approved, attempted, confirmed, or failed. These are design suggestions, not a universal standard. Their purpose is to prevent a generic “done” message from concealing uncertainty about the actual outcome.
Retries should account for side effects. Repeating a read is different from repeating a purchase or sending the same email twice. Where the surrounding system supports it, use an operation identifier to recognize duplicate attempts. Otherwise, inspect the state before trying a consequential action again.
Measure the whole loop
Our recommended evaluation asks whether the agent completed the intended task within its boundaries. Record incorrect actions, incomplete work, unnecessary reviews, and the time a person spent correcting the result. An assistant that produces impressive drafts but transfers substantial checking work to the user may deliver less value than its demo suggests.
Include cases with missing information, conflicting instructions, and unavailable tools. These reveal whether the system stops cleanly or invents a way forward. Keep evaluation inputs separate from examples used to improve the prompts.
A good first deployment is a bounded task with clear evidence of success and a practical recovery path. Expand the agent’s authority only when evaluation supports doing so. Autonomy is a design choice to justify, not a score to maximize.
Sources & further reading
- Anthropic — Building effective agents
Foundational engineering essay, originally published December 2024. Architectural guidance; product examples change over time.
Sources checked 5 October 2026. Links support the specific research claims identified in the article; recommendations and examples are ReasoningReport analysis.
