iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Reliable AI agents depend on more than the model. In his August 25, 2026, Gradient Flow article, Ben Lorica argues that teams should evaluate and design the whole system: its tools, context, memory, policies, workflow, and recovery logic. His nine rules offer a practical way to reduce avoidable errors while keeping autonomy focused on work that genuinely needs it.
1. Enforce hard constraints in software
Use a model where judgment is ambiguous; use software and policy controls where behavior must be predictable. Calculations, permissions, state changes, retries, code validation, and factual checks are better handled by mechanisms that can enforce rules consistently.
For consequential actions, do not rely solely on a prompt telling the agent what not to do. Design the system so a prohibited action is blocked or requires approval. A prompt can guide behavior, but it is not the same as an access control.
2. Give the agent only the autonomy the job needs
More autonomy means more possible paths, opportunities for error, cost, and governance work. Start with enough flexibility to learn which steps are useful, then move stable, repeatable paths into ordinary code. Lorica’s principle is that “A mature agent faces fewer open-ended choices over time, not more.”
#1 Best Overall
This is not a call to eliminate judgment. Keep the model involved where circumstances are variable or interpretation matters; constrain or automate the parts that have become routine.
3. Structure work around a trusted domain process
If a field already has a proven checklist, protocol, or decision process, use it as the agent’s workflow rather than inventing a free-form alternative. A support triage process, for example, can specify what information to collect, when to escalate, and which actions need approval. Medical diagnostic protocols are another example of processes that define information-gathering and escalation steps.
The process gives the agent a reliable sequence to follow and makes it clearer where human review belongs.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →4. Design for recovery, not a perfect first attempt
In a long workflow, small per-step error rates can compound. Lorica illustrates this with a calculation: if each of ten steps succeeds 95% of the time, and step outcomes are independent, the chance of an entirely error-free sequence is about 60%. This is an illustrative calculation, not a measured result for a particular agent.
Build for the possibility that a step fails or produces an unintended result:
- Add checkpoints so the system can identify what has already succeeded.
- Verify that an action achieved its intended effect instead of assuming it did.
- Provide retries where they are safe, and make actions reversible where possible.
- Allow work to resume from a known-good state.
- Measure recovery separately from first-attempt accuracy.
5. Evaluate the model and its harness together
An agent is not just a model. Its harness includes the tools it can call, context management, memory, policies, and recovery logic. Evaluation should use the actual tools, permissions, context, and failure modes the system will encounter—not a simplified setup that omits the operating environment.
When either the model or the harness changes, rerun the evaluations. Lorica reports an 18-percentage-point difference between the best and worst harness configurations for the same open model. The article does not identify the model, evaluation, or underlying report, so the figure should be read as a result he reports, not as a general benchmark.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →6. Keep multi-agent teams small and accountable
Multiple agents make sense when work is genuinely parallel or when independent perspectives are useful. Otherwise, coordination can create duplication and make failures harder to diagnose. Lorica recommends a clear orchestrator and a small number of specialists with distinct roles, tools, permissions, and information.
Give a critic an explicit standard and the authority to block or escalate work that fails it. A nominal reviewer without defined criteria or decision power may add coordination without providing a meaningful check.
7. Keep the toolbox focused
Overlapping tools make selection harder, expand the number of possible call sequences, and can increase the systems an agent is able to read or change. Combine, route, or remove tools that do the same job. Log which tools are selected, their inputs and outputs, and any failures so the team can see how the agent actually uses them.
8. Separate context, memory, and enterprise knowledge
These three kinds of information serve different purposes, even if they share infrastructure:
- Context is the information the model needs for the current run.
- Memory is information worth retaining from past work.
- Enterprise knowledge is the governed set of records, documents, and policies available for retrieval.
Give each suitable retention, retrieval, and access rules. Memory should preserve useful lessons, not indiscriminately store transcripts. Enterprise knowledge, meanwhile, needs governance appropriate to the records and policies it contains.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.9. Improve knowledge before reaching for a bigger model
When an agent gives poor answers, first check whether it receives the right information in a usable form, with appropriate structure, routing, and governance. Retrieval can miss relevant material when a user’s wording differs from the documents, key details are buried in tables or PDFs, or sources conflict.
Lorica describes one support team that replaced raw documents with a diagnostic playbook routing complaints to symptoms, problem types, and handbook material. He reports 43% fewer tokens and 48% fewer errors without changing the model. The team and underlying measurements are not identified in the article, so these figures describe that reported example rather than an independently established result.
What to evaluate before deployment
Use the nine rules as a review of the whole operating system, not as a prompt-writing checklist. Before deployment, check that the system has:
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute- Software-enforced constraints and permissions for actions that must not be left to model judgment.
- An autonomy level matched to the task, with stable paths moved into code where appropriate.
- A domain workflow that defines information gathering, escalation, and approval.
- Checkpoints, verification, safe retries, reversibility, and a way to resume after failure.
- Evaluations that reflect the actual harness, tools, context, permissions, and failure conditions.
- A small, clearly coordinated agent team if multiple agents are needed, including an empowered critic.
- A focused toolset with logging of selection, inputs, outputs, and failures.
- Distinct governance for current context, retained memory, and enterprise knowledge.
- A knowledge-quality review before treating a larger model as the default fix.
Lorica’s broader point is that model choice is only one part of reliability: “The model is the component you can buy. The system around it is what you own.” Read the original article in Gradient Flow.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
