Reliable AI agents depend on more than a capable model. They need software-enforced limits, appropriately scoped autonomy, a workflow designed for the domain, and a way to detect and recover from mistakes. Ben Lorica’s nine practical rules for agents doing real work offer a useful design framework—not a universal standard—for building and evaluating that whole system.
1. Put hard constraints in software
Use the model for ambiguity and judgment, not as the sole mechanism for enforcing permissions, calculations, or predictable control flow. A prompt can tell an agent not to take an action, but it does not reliably prevent that action. As Ben Lorica puts it, “A prompt is guidance.”
As an Amazon Associate I earn from qualifying purchases.
Enforce consequential limits in code or policy controls. Validate the agent’s output and check critical factual claims before they drive an action. For example, an agent may draft a reimbursement decision, but software should enforce spending limits and require an authorized approval when policy calls for one.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →2. Give autonomy only where the job needs it
More autonomy creates more possible action paths, and therefore more ways for a workflow to fail. It can also increase cost and the burden of oversight. Start with the narrowest authority that lets the agent do useful work; grant additional permissions only when the task requires them.
#1 Best Overall
As a workflow becomes well understood, move repeatable, reliable steps into ordinary code. The agent can then focus on decisions that genuinely need interpretation rather than improvising through steps that can be handled predictably.
3. Build around the domain’s trusted process
Do not assume a generic plan-and-act loop is the right workflow for every job. Where the domain already has checklists, protocols, escalation rules, or approval points, make those the structure the agent follows.
This keeps the agent inside recognizable procedures and makes it clearer when a case needs human attention. A useful design question is not only “Can the agent complete this task?” but also “What established process governs this task, and where does that process require a check or escalation?”
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware match4. Design for recovery, not just a successful first attempt
Long workflows can fail midway, even when each step usually succeeds. Lorica illustrates the compounding effect with an example: at a 95% success rate per step across ten independent steps, the chance of an error-free run is about 60%. This is an author-reported illustration, not an independently verified benchmark.
Make failure manageable with checkpoints, post-action verification, retries where safe, reversible actions, and the ability to resume from a known-good state. Measure recovery separately from first-attempt accuracy: a system that detects and corrects a mistake may be more dependable in practice than one that performs well only when nothing goes wrong.
5. Evaluate the model and its harness as one system
When you test an agent, evaluate the complete system that performs the work—not just the underlying model. The harness includes its tools, context management, memory, policies, and recovery logic. A strong model can still produce an unreliable workflow if these surrounding pieces are poorly designed.
Repeat evaluations when either the model or the harness changes. Lorica reports an 18-percentage-point gap between the best and worst harness configurations for the same open model. The article does not provide the underlying study’s methods or sample, so treat this as an example reported by Lorica rather than 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 make criticism consequential
Use multiple agents when distinct responsibilities justify them, not simply because coordination is possible. Give agents separate roles, tool sets, permissions, and only the information they need. This limits unnecessary overlap and makes it easier to understand which part of a workflow is responsible for an outcome.
If one agent reviews another’s work, define what the reviewer should check and give it explicit authority to block an action or escalate a case. A critic that can only offer optional feedback is not a meaningful safety check.
7. Keep the toolbox compact and distinct
Overlapping tools can make it harder for an agent to choose the right one. They also expand the number of tool-call sequences that need to be tested. Prefer a compact set of tools with clearly differentiated purposes; combine, route, or remove tools when their responsibilities duplicate one another.
Log tool selections, inputs, outputs, and failures. Those records help reveal whether a tool is being chosen incorrectly, whether its result is misunderstood, or whether the tool itself is failing.
Recommended Free Tools
8. Separate context, memory, and enterprise knowledge
These three information sources serve different purposes, so they should not be treated as one undifferentiated store.
Best Value
- Context is the information needed for the current run.
- Memory carries relevant lessons forward between runs.
- Enterprise knowledge is governed material the system may consult.
Set retention, retrieval, and access rules to match each role. In particular, governed organizational material should have access controls appropriate to its sensitivity; information useful in one run does not automatically belong in persistent memory.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.9. Improve knowledge before paying for a larger model
When an agent gives an answer that misses relevant information, first check whether the knowledge workflow is the problem. Information can be difficult to retrieve when a user’s wording differs from the source, key details are buried in tables or PDFs, or documents conflict.
Examine document structure, routing, and governance before changing models or fine-tuning. Lorica reports an example in which replacing raw support documents with a diagnostic playbook and routing approach reduced token use by 43% and errors by 48%, without changing the model. The article does not identify the underlying study’s methods, sample, or organizations, so these are figures reported by the author, not independently established performance guarantees.
Free tools Windows power users keep installed
One-click scans. No signup required.
Turn the rules into an evaluation plan
For a particular workflow, use these questions to check whether the design and its evaluation cover the important failure modes:
Quick Recap
- Which actions or outcomes must be enforced by software rather than requested in a prompt?
- What level of autonomy does the job require, and which repeatable steps can become ordinary code?
- What domain procedures, approvals, and escalation rules should shape the workflow?
- Can the system verify consequential actions, recover from failure, and resume from a known-good checkpoint?
- Does evaluation cover the model, tools, context, memory, policies, and recovery logic together?
- Are agents’ roles and permissions distinct, and can a reviewer actually block or escalate?
- Are tools clearly differentiated, and are selections and failures logged?
- Are current-run context, persistent memory, and governed knowledge separated by appropriate access and retention rules?
- Could document structure, routing, or conflicting sources explain a knowledge failure before a model change is considered?
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.
