Harness Safe Defaults for AI Agents: What to Set Before Launch
Learn which safe defaults a harness should ship with before an AI agent goes live: read-only access, rate caps, disabled outbound messages and approval gates.

A safe default is a setting a harness ships with before any agent receives live work: read-only access, low rate and step caps, outbound messages turned off, and an approval step for anything that cannot be undone. Each default should loosen only when evidence from draft runs shows the agent does what its task contract says. The sections below list the defaults to set first and the order in which to set them.
Key takeaways
- Start every agent with read-only access to the systems it needs, and grant write access one named action at a time.
- Cap calls, steps and spend per task so that a loop fails small instead of large.
- Keep outbound email, messages and posts disabled until a reviewer has approved sample output.
- Put approval gates on any action that cannot be reversed, and write the stop conditions down before launch.
- Store every default in a versioned config so that each later change is visible and reversible.
Why defaults matter more than prompts
Many teams spend days on a system prompt and leave the harness settings at whatever the framework shipped with. That order is backwards. The prompt describes what the agent should try to do. The harness decides what the agent can reach, how often it can act, and what happens when something goes wrong. A careful prompt does not stop an agent with write access to a shared folder from overwriting the wrong files during a loop it cannot see from inside the task.
Untrusted input adds a second reason. Emails, web pages and uploaded documents can contain text written to steer an agent toward actions its owner never intended. The OWASP guidance on prompt injection prevention assumes that some input will be manipulative and focuses on limiting what the model can do when that happens. Defaults are that limit, so they should be in place before the first real input arrives, not added after an incident.
Which access should an agent start with?
Start from the narrowest access that lets the agent finish its first task, then widen it one permission at a time. A practical order looks like this:
- Read-only access to the specific records, folders or queues the task uses, with no broad view of the whole system.
- Write access switched off by default, then enabled for one named action with a named reviewer.
- A dedicated service credential for the agent, scoped to one workspace, never a shared administrator login.
- A tool allowlist that contains only the functions the first task requires, reviewed before each widening.
Scoping is not a one-time setup. The OWASP authorization guidance points toward checking permissions on each request rather than assuming that access granted at setup still applies. For an agent, every tool call should be evaluated against the current allowlist and credential scope, so a change made earlier does not silently extend what the agent can do today. The guide to tool allowlists for AI agents explains how to widen that list using evidence.
How should rate caps, step limits and spend ceilings be set?
Rate caps limit how many calls an agent may make to a tool in a given period. Step limits cap how many planning and action cycles a single task may consume. Spend ceilings cap model usage per task and per day. All three are cheap to configure and costly to lack, because a loop that repeats a failing call can touch many records before anyone notices. Set them low for the first run, and raise a cap only when logs show that legitimate tasks regularly reach it. The article on token and step budgets shows how to pair these caps with alerts.
Set these defaults before the first run:
- A call limit for each tool per hour, low enough that a looping task stops before it reaches many records.
- A step limit for each task, with the task failing closed and reporting its state when the limit is reached.
- A spend ceiling per task and per day, with an alert that fires well before the ceiling.
- A retry limit for failed tool calls, so the same error is not repeated indefinitely.
Test each cap on purpose. Give the agent a task that cannot succeed and confirm that it stops at the limit, records the reason and does not retry past it. A cap that has never fired in testing is a cap you have not verified.
Should outbound messages be disabled at launch?
Yes, by default. Outbound messages are the action most likely to cause harm that cannot be recalled: an email sent to the wrong customer, a chat reply that states a policy the business does not have, or a post that publishes the wrong price. Launch in draft mode so the agent writes every message into a review queue. Enable sending for one channel and one message type at a time, and only after reviewers have approved sample drafts against the task contract.
Illustrative scenario: a small clinic gives a scheduling agent the job of sending appointment reminders. In the first days it only drafts reminders into a queue that the front desk lead reviews. Once several batches show correct times and wording, the team turns on sending for standard reminders only. Cancellations, billing questions and anything that mentions results stay in the queue for a person to answer. Each step is recorded in the config with the name of the person who approved it.
Stop conditions belong in the same decision. If the agent meets a situation it was not built for, such as a message from an unknown sender or a tool that keeps failing, it should halt and escalate rather than improvise. The article on stop conditions for agent loops describes how to define those triggers.
A launch checklist for harness defaults
Run this checklist before the first live task. Each item needs an owner and a written record, even when one person owns several items.
- Write the default configuration in one file and version it before any credential is issued.
- Confirm that every tool is on the allowlist and that every write action names an approver.
- Switch every outbound channel to draft-only mode.
- Set call, step, retry and spend caps, then test that each one stops a deliberately looping task.
- Define the stop conditions and name the person who receives each escalation.
- Run the agent against sample inputs, including messages that try to override its instructions.
- Record what was tested, what failed, and what changed as a result.
- Schedule the first permission review for a fixed date after launch.
Run these tests in a copy of the systems rather than the live ones. The difference between the two environments is explained in the guide to staging and production for AI agents.
How do defaults stay in place after launch?
Defaults usually drift for an ordinary reason: a bug appears late on a Friday, someone widens a permission to fix it, and the wider setting stays long after the bug is gone. Prevent this by keeping the configuration in version control, requiring a short note with every change, and reviewing permissions on a schedule. The NIST AI Risk Management Framework treats governance and monitoring as continuing activities across the life of a system, not a step completed once. Anthropic's engineering guidance on long-running agents makes a similar point from the builder's side, describing how progress and state should be written down so that each session can be checked against a known record.
Be Human designs, implements and operates personalized AI infrastructure for companies, including the harness that holds an agent's identity, memory, channels, tools and permissions. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion, and the defaults above are written into that harness before the agent handles live work. Managed operations then review those settings, monitor the agent, run recurring evaluations and respond to incidents when a stop condition is reached.
If you are not sure which task should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Are safe defaults the same as guardrails? Not quite. Guardrails are checks on inputs and outputs, such as blocking certain content or validating a field. Safe defaults are the starting settings for access, limits and channels that decide what an agent can reach at all. A business usually needs both, but defaults come first because they limit the damage a guardrail failure can cause.
Should we wait until the model is proven before setting defaults? No. Defaults matter most at the start, when nothing has been proven yet. A conservative harness lets you learn how the agent behaves without exposing customers or live records. Proof comes from draft runs under those defaults, and every loosening should point to the evidence that justified it.
Who should approve a change to the defaults? Assign one owner for the harness configuration and a named approver for each category of change, such as new write access or a new outbound channel. The owner should be someone who can answer for the business risk, not only the technical setup. Record each approval with its reason and the test that supported it.
Can a small team set these defaults without an engineering department? Often yes, provided someone can edit a versioned configuration file and read the logs. The main requirement is discipline rather than headcount: the team must decide which actions are irreversible and refuse to skip review for convenience. If nobody can read logs or rotate credentials safely, a managed operations partner can carry that work.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


