# Relevance AI Agents: How They Decide, Pause and Bill Each Run

> How Relevance AI agents work: models, tools, approval modes, autonomy limit, memory and triggers, and why every tool run, failed or not, costs one Action.

_Source: https://professionalstoolkit.com/articles/relevance-ai-agents — The Professional's Toolkit · updated November 2026_

---

A Relevance AI agent is a language model that gets a set of tools and decides for itself which one to run, and in what order, to reach a goal. Each tool can run on its own, wait for approval, or leave that call to the agent, and an Agent Autonomy Limit caps tool runs in a row before someone signs off. Every tool run is billed as one Action [1], failed runs included [2]. Plan around that number. Agents suit judgment-heavy work with a human in the loop; fixed processes belong inside a tool.


**1 Action** — Every tool run an agent makes costs one Action, including failed runs and retries.



- **3** — Approval modes per tool: Auto Run, Approval Required, Let Agent Decide
- **1 retry** — What a fallback model gives a failed task
- **Enterprise** — The only plan with agent evals


## What an agent is built from

Every agent starts from a prompt: the instructions that set the job and the behavior. Invent, the built-in AI builder, can write a structured prompt from a few words, though the tools it invents may still need finishing in the separate Tool Builder. Then come the model and the tools. Each tool input can be left for the agent to fill, set by hand, or taken from an earlier tool's output. Setting an input by hand is how you stop an agent improvising values that should never change, such as a pipeline stage or a sender address; taking it from an earlier tool's output passes results along without the agent retyping them.

Knowledge attaches in one of two ways. A small reference set can be added in full to the prompt on every run, while a large one is searched by the agent with retrieval, using semantic, hybrid or metadata-filtered search across one or several knowledge bases. Variables store a value once for reuse in prompts and tools, and snippets are named text blocks you edit once to update everywhere. Two agent types sit higher up: phone agents that place outbound calls and meeting agents that join Google Meet, Zoom or Teams both start on Team.

## Picking the model, and what happens when it fails

You can leave the model to Relevance AI, which offers cost-optimized and performance-optimized defaults, or pick one from OpenAI, Anthropic Claude, Azure OpenAI, Google Gemini and OpenRouter. Not every model is eligible. An agent needs strong tool calling, so the docs suggest keeping reasoning models such as DeepSeek inside an LLM step of a tool instead of the agent seat.

Three settings shape reliability and cost. A fallback model retries a failed task once, and the docs advise taking it from a different provider than the primary, so one vendor's bad hour does not sink both attempts. Some Claude models add a thinking type and an effort level, but reasoning tokens bill as output tokens, so deeper thinking costs more. Parallel Tool Calls, still in beta, lets an agent run several tools at once. Your own model keys are optional on every plan and cut Vendor Credit spend, but not the subscription or Actions.


> 💡 **Fallback from another provider:** The docs advise a fallback model from a different provider than the primary, so one provider's outage does not fail both attempts.


## How much freedom the agent gets

Freedom is set per tool. Each tool runs in one of three approval modes: Auto Run, Approval Required, where the agent asks permission first, or Let Agent Decide [3]. The docs' Outlook example starts sending tools in approval mode and moves them to autopilot once the drafts prove reliable. Pending items land on the Tasks page, which has tabs for Approvals, Escalated, Errors and All, with bulk approve and reject. Since January 2026, tool requests can also be approved or rejected from Slack.

The Agent Autonomy Limit is the second brake. It caps consecutive tool runs without a human approving, and it counts every run, retries included [4]. If Tool A runs three times and Tool B seven times, the agent has used ten runs of its limit. That double counting matters. A loop that keeps retrying a broken tool eats the autonomy budget and the plan's Actions at the same time, so the failure policy deserves a deliberate choice.

| Control | What it limits | What happens at the limit |
|---|---|---|
| Approval mode (per tool) | Whether a tool runs unattended | The agent waits for a person to approve or reject |
| Agent Autonomy Limit | Consecutive tool runs, retries included | A human approval is needed to continue |
| Failure policy | Forced retries of an errored tool | The task ends or escalates to a person |
| Fallback model | A failed task | One retry on the second model, then the task fails |
| Escalation | Questions the agent cannot resolve | A person answers by email or chat |

The failure policy forces retries of an errored tool up to a maximum count, then either ends the task or hands it to a person, and the docs warn that more retries can raise credit usage. Escalation is the third brake. An agent that cannot resolve something passes it to a human by email or chat, and the answer can be stored for reuse. An Escalate to Manager tool, switched on from the Alerts page, hands a conversation the agent cannot handle to a human manager.

Relevance AI's guides describe four autonomy levels, from L1 Assisted and L2 Copilot to L3 Autopilot and L4 Self-driving, and these controls are how you move between them. Keep one property in mind. The troubleshooting docs say agents make their own decisions and are not end-to-end workflows, so they may not take the exact path you expected [5]. When steps must run identically every time, put them inside one tool, and when a job needs several specialists, the docs recommend a [Workforce](/articles/relevance-ai-workforce) over the legacy sub-agents.

## Memory and the triggers that wake an agent

Short-term memory, which the docs also call metadata, lasts one conversation or run. Its fields take five formats (text, number, true/false, single option and multiple option) and can be filled by the agent or by rules that fire after a tool run or a trigger. Long-term memory persists across conversations and adds two tools, Save to Memory and Delete from Memory. You can switch off the delete tool so the agent cannot erase what it stored. Deleting a memory field yourself is permanent.

Agents start from a schedule, an outside event or a manual run. Schedules come as Daily, Weekly, Monthly or Custom dates, or as a cron expression in the AWS EventBridge format, which has six fields including the year. More than 1,000 app triggers are available, and the build-your-own options are a webhook, the API trigger, the JavaScript SDK and tools used as triggers, and a webhook can be deduplicated with a Unique ID.


> 💡 **Map a Unique ID on webhooks:** With a Unique ID field mapped on a webhook trigger, a repeated event is ignored, so the same lead or ticket is not processed twice.


## Clone a template, or describe the job to Invent

There are three ways to get an agent: describe it to Invent, clone one from the Marketplace, or build it from scratch with a prompt, tools and optional knowledge. The Marketplace groups agents, tools and Workforces under Sales, Marketing, Content Creation, Research, Operations, and HR and talent, with templates such as Sales Researcher and Outbound Composer. A Clone Agent button copies the agent, and you then connect the channels it needs. Our [step-by-step tutorial](/articles/relevance-ai-tutorial) walks through a first build.

Cloning copies less than it looks. Knowledge tables, triggers (only suggested), global snippets, secrets and API key or OAuth values stay behind, so a clone needs its data and credentials reconnected before it runs. Running any cloned agent also needs a Pro or Team subscription, and a paid listing costs extra. The named agents on Relevance AI's use-case pages, such as the Ticket Resolver, and the BDR agent (earlier marketed as Bosh) are different: sales-led engagements booked through a demo and tuned by Relevance AI's experts, not templates you clone.
## What every agent run costs

Two meters run under each agent. Actions count tool runs: one email or a many-step workflow inside one tool is one Action. Vendor Credits cover model and tool costs at wholesale with no markup, at $20 per 10,000 credits. By the docs' own account, running out of either is the most common reason an agent fails, so set usage alerts and limits, which every plan has, before an agent goes on autopilot. The vendor's tip is to combine steps into one tool to save Actions; our [tools guide](/articles/relevance-ai-tools) shows how.

Plans set the budget. Pro includes 2,500 Actions a month and Team includes 7,000 Actions a month. Beyond the allowance, extra Actions come as a top-up pack at $80 per 1,000 Actions, or $0.08 an Action by our arithmetic, and purchased Action top-ups roll over to the next billing cycle. Retries are where budgets leak: a tool that fails and is retried three times before escalating spends four Actions on one failure, by our arithmetic. Every organization now needs a paid plan, and Relevance AI's pricing pages describe no free trial.


> ⚠ **Retries spend Actions:** Every retry of a Relevance AI tool is another tool run, so it counts toward both the Agent Autonomy Limit and the plan's Actions.


Testing has its own gate. Evals, the simulated test conversations and pass-or-fail checks that can block a weak version from publishing, are available only on Enterprise, and each check costs one Action plus Vendor Credits. On Pro and Team, the safety net is approval modes plus version history: every draft save and publish is versioned, and restoring puts the old version back into draft. Watch sharing as well. An embedded or cloneable agent is public, and anyone with the link runs it on your credits.


## FAQ

**Can I export a Relevance AI agent?**

Yes. An agent's configuration can be exported as a .rai file in JSON format, for backups, sharing, moving between environments or reuse as a template.

**Can a deleted agent be recovered?**

No. Deleting a Relevance AI agent is permanent, and Relevance AI support cannot recover it, so export a .rai backup first.

**Can I build or use agents on a phone?**

You can use agents through the Relevance Chat apps for iOS and Android, but the builder runs only in a desktop browser.

**Can a Relevance AI agent call tools on other MCP servers?**

Yes, through the MCP client, which reaches remote servers only over Streamable HTTP; Notion, Canva and Atlassian are offered as presets.


## Before an agent runs without you watching

Start each new agent with its sending and writing tools on Approval Required, a modest Autonomy Limit and a failure policy that escalates after a retry or two instead of looping. Attach knowledge by search unless the set is small, map a Unique ID on any webhook trigger, and set a usage alert before the first scheduled run. Then read the Approvals and Errors tabs for a couple of weeks. When the drafts you approve stop needing edits, move that one tool to Auto Run, and keep the rest supervised until they earn the same trust.



## Sources

[1] Relevance AI documentation: Pricing (plans, Actions, Vendor Credits) — https://relevanceai.com/docs/get-started/pricing (2026-11)
[2] Relevance AI documentation: Plans and credits — https://relevanceai.com/docs/admin/subscriptions/plans (2026-11)
[3] Relevance AI documentation: Give Your Agent Tools — https://relevanceai.com/docs/build/agents/build-your-agent/tools (2026-11)
[4] Relevance AI documentation: General — https://relevanceai.com/docs/build/agents/build-your-agent/agent-settings/general (2026-11)
[5] Relevance AI documentation: Troubleshooting agents not working — https://relevanceai.com/docs/get-started/troubleshooting/troubleshooting-agents-not-working (2026-11)
[6] Relevance AI Terms and Conditions — https://relevanceai.com/terms-and-conditions
[7] Relevance AI documentation: Changelog — https://relevanceai.com/docs/changelog