When AI summarizes a document, a person can review the result before using it. When AI updates customer records from that document, mistakes may enter the business directly. Similar model outputs have different consequences and therefore need different architectures.
Moving from conversation to execution adds permissions, state, verification, and recovery—not merely tool interfaces. Reliable systems place model judgment inside observable, constrained processes.
Let the Task Determine Autonomy
Anthropic distinguishes workflows with predefined code paths from agents that dynamically choose processes and tools. Its engineering guidance favors simple solutions and additional complexity only when useful. This is practical guidance, rather than a mandatory architecture for every application. Building effective agents.
Invoice entry can follow fixed steps: read, extract, validate amounts, and save. Changes in document fields need not justify an autonomous planner. Researching an unfamiliar supplier may need adaptive searches, but searching and signing a contract should have different permissions.
Evaluate autonomy against task variability, error cost, and verifiability. Clear paths and rules favor deterministic workflows. When paths cannot be enumerated, give the model a bounded action space.
Give Context a Source and Scope
Longer context does not automatically mean better results. Old quotations, draft discussions, and current contracts can compete for a model's attention.
The system should identify where information came from, when it applies, which customer it concerns, and whether it remains unconfirmed. Refund policies, order facts, and historical staff opinions need separate labels. An old conversation must not silently become current authorization.
Memory also creates maintenance, retrieval, permission, and expiry work. Its value lies in reducing mistakes or repeated effort, rather than in accumulation itself.
Separate Recommendations from State Changes
A model may recommend a refund. The transaction system must execute it and report the actual outcome. A statement of success is not proof of execution.
Interfaces should check identity, amounts, resource scope, and business prerequisites. Repeated submissions need idempotency; failures need inspectable state. External messages, transfers, and deletions need confirmation or approval proportional to their consequences. A prompt asking the model to be careful is insufficient.
Useful traces record sources, tool parameters, outcomes, and failure reasons. Logging does not justify storing all sensitive information indefinitely; access and retention require their own limits.
Verify During Development and Operation
Conventional software also needs production monitoring, access controls, and runtime validation. AI adds outputs and paths that can vary with input, including plausible-looking errors. Both pre-release checks and operational sampling matter.
A meeting summary can be checked for names, dates, decision evidence, and omissions. A refund workflow additionally needs real transaction state, correct amounts, and duplicate detection. Programs can check format and rules; human or model assessment can assist with complex quality judgments, provided standards are calibrated.
Anthropic's discussion of agent evaluations covers graders, traces, and evaluation environments. The principle used here is to inspect final outcomes and investigate failed trajectories, rather than allowing one aggregate score to substitute for all risks. Agent evaluation guidance.
Treat Budget as an Operating Constraint
Autonomous loops can keep searching, retrying, and calling tools. Small costs per step may accumulate.
Set limits on calls, elapsed time, and spending per task, with a pause or escalation when limits are reached. Efficiency should use accepted results as the denominator. Include unsuccessful runs and human intervention in the numerator.
A more expensive model can be cheaper overall if it reduces rework. A cheaper model that reliably meets the standard within a constrained task needs no upgrade merely for prestige.
Make One Workflow Reliable First
For customer-record updates, a first version can read documents and propose sourced changes for staff approval. A later version can check duplicates and required fields in code, then allow limited low-risk automatic saves. Error rates, intervention, and cost should govern further expansion.
There is no required destination called “sufficiently agentic.” More autonomy is worthwhile when it improves delivery within clear boundaries. Connecting goals, information, authority, execution state, verification, and recovery says more about maturity than the number of agents.
Sources checked through October 1, 2026. Business workflows are design examples. Architectural suggestions do not certify any application as safe or effective.