Reliable AI agents depend on more than the model. In Ben Lorica’s August 25, 2026, article for Gradient Flow, the central recommendation is to evaluate and design the whole system: its tools, context, memory, policies, recovery logic, and the workflow and knowledge it receives. The nine rules below turn that principle into practical design choices.
1. Put hard constraints in software
Use a model where a task requires judgment; use ordinary code and policy controls where the result must be predictable. A prompt can guide an agent, but it should not be the only safeguard for consequential actions.
As an Amazon Associate I earn from qualifying purchases.
- Enforce permissions and access boundaries outside the prompt.
- Use code for calculations, predictable state changes, retries, and validation.
- For actions that should not happen without oversight, prevent them in the system or require approval.
This division makes the model responsible for ambiguity without asking it to serve as the final security or correctness control.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitches2. Limit autonomy to what the job needs
Every open-ended choice creates another route the agent might take—and another opportunity for error, cost, or governance work. Start with enough flexibility to learn what the task requires. When a path proves stable and repeatable, move that part into conventional code.
#1 Best Overall
Lorica’s formulation is useful: “A mature agent faces fewer open-ended choices over time, not more.” Autonomy is a design setting to revisit, not a goal to maximize.
3. Use the domain’s trusted process
If a field already has a proven checklist, protocol, or decision process, use it to structure the agent’s work. Lorica points to support triage and medical diagnostic protocols: both can specify what information to gather, when to escalate, and where approval is needed.
Rather than asking an agent to improvise an entire workflow, encode the established sequence and let the model help with the judgment calls within it.
4. Design for recovery, not a perfect first attempt
Small error probabilities compound across long workflows. Lorica illustrates this with a hypothetical sequence: if each of ten steps succeeds 95% of the time, and step outcomes are independent, the chance that all ten succeed is about 60%. This is an illustrative calculation, not a measured agent result.
Build workflows so that errors can be found and contained:
- Add checkpoints at meaningful transitions.
- Verify that an action produced its intended result.
- Provide retries and reversible actions where possible.
- Allow work to resume from a known-good state.
- Measure recovery separately from first-attempt accuracy.
5. Evaluate the model and harness together
When you ask, “What exactly are you evaluating when you test an agent?”, include the harness: the tools, context management, memory, policies, and recovery logic around the model. A model score alone cannot tell you whether the deployed system works with its actual permissions, information, and failure modes.
Test with the tools and operating conditions the agent will really use. Rerun evaluations when you change either the model or the harness. Lorica reports an 18-percentage-point gap between the best and worst harness configurations for the same open model, but his article does not identify the model, evaluation, or underlying report; treat that figure as an attributed report, not a general benchmark.
Recommended Free Tools
6. Keep multi-agent teams small and accountable
Use a clear orchestrator and a small set of specialists with distinct roles, tools, permissions, and information. More agents can help when work is genuinely parallel or when independent perspectives are useful. Otherwise, coordination can create duplicated effort and make failures harder to diagnose.
If a critic agent is part of the system, give it explicit criteria and authority to block or escalate. A reviewer without a defined standard or an effective way to intervene is not a dependable safeguard.
Rank #4
7. Keep the toolbox focused
Overlapping tools make it harder for an agent to choose the right one, expand the number of possible call sequences, and may broaden the systems it can read or change. Combine tools, route requests deliberately, or remove tools whose jobs overlap.
Log tool selection, inputs, outputs, and failures. Those records help reveal whether a tool is poorly scoped, incorrectly selected, or failing after selection.
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 →8. Separate context, memory, and enterprise knowledge
These are related but different information layers, and they need suitable rules for retention, retrieval, and access.
Best Value
- Context is what the model needs for the current run.
- Memory is information worth carrying forward from past work.
- Enterprise knowledge is the governed collection of records, documents, and policies available for retrieval.
They may share infrastructure, but they should not be treated as interchangeable. Memory should retain useful lessons rather than indiscriminately keeping transcripts; enterprise knowledge needs governance suited to its source material and users.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.9. Improve knowledge before changing models
Before changing a model or fine-tuning it, diagnose whether it receives the right information in a usable structure, with appropriate routing and governance. As Lorica puts it, “Before changing the model, improve the structure, routing, and governance of the knowledge it receives.”
His article describes one unidentified team that replaced raw support documents with a diagnostic playbook mapping complaints to symptoms, problem types, and handbook material. The team reportedly used 43% fewer tokens and had 48% fewer errors without changing the model. Those figures are claims reported by Lorica; the article does not name the team or link an underlying measurement.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Retrieval can still fail when a user describes a problem differently from the documents, when useful information is buried in tables or PDFs, or when sources conflict. A playbook and routing structure can help, but they do not remove the need to check whether retrieved information is relevant and trustworthy.
What these rules change about agent design
Taken together, the recommendations move the engineering question away from “Which model should we use?” and toward “What system makes this task dependable?” The model remains important, but constraints, workflow, tools, information quality, and recovery determine what it can safely accomplish in practice. Lorica’s summary is: “The model is the component you can buy. The system around it is what you own.”
Quick Recap
Read Ben Lorica’s full article in Gradient Flow.
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.
