> ## Documentation Index
> Fetch the complete documentation index at: https://docs.gumloop.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Agents
Agents are AI-powered assistants that use tools to solve open-ended tasks. An agent decides which tools to use and when, adapting its approach to the task in front of it.
## What agents do
You give an agent a goal and a set of tools, and it figures out how to get there: which tools to call, in what order, and when to ask you for input.
* **Adaptive**: takes different approaches for different situations
* **Tool-driven**: uses your connected services and data as needed
* **Conversational**: works through a task in a back-and-forth conversation
* **Context-aware**: considers your instructions, skills, and past tasks
One piece of work you hand to an agent is a **task**. Each task is its own conversation with its own history, files, and credit cost.
When you give it a task, the agent analyzes the request, decides which tools to use and in what order, runs them, adapts based on the results, and asks for confirmation when your instructions call for it.
## Configure an agent in three moves
You do not need to write a long system prompt up front. The fastest way to a useful agent is to let it do real work, then have it write its own instructions.
Open the **Connectors** section and add the integrations the agent needs (Gmail, Salesforce, Slack, and so on). The connectors you add define what the agent can see and do, so start with the two or three it actually needs.
Start a task and run something end to end. Watch where it guesses, asks the wrong thing, or misses a step.
Once it completes a task the way you want, tell it: *"Update your instructions so you always do it this way"* or *"Turn this into a skill."* The agent writes its own [instructions](#self-improving-instructions) and [skills](#skills), so you can capture the approach without writing everything from scratch.
When an agent gets something wrong, give it feedback and ask it to save the correction in its instructions or a skill. Saving the change gives future tasks a reusable starting point; feedback in one conversation alone does not update every future task.
## Getting around an agent
Open an agent and the left sidebar has everything about it:
| Item | What it is |
| - | - |
| **New Task** | Start a fresh task. `Ctrl/Cmd + Shift + O`. |
| **Artifacts** | Every file, document, and chart the agent has produced. See [Agent Artifacts](/core-concepts/agent_artifacts). |
| **Performance** | Insights, Evaluations, and Tasks for this agent. See [Agent Performance](/core-concepts/agent_performance). |
| **Channels** | Where the agent is reachable: email, Slack, Microsoft Teams, a hosted page, the API, and MCP. See [Channels](#channels). |
| **Recents** | Your recent tasks, newest first, with a filter. |
When the agent is waiting on you to approve a tool call, a **pending action request** appears above Recents.
## The agent builder
Everything you configure lives in four tabs at the top of the builder. Only Owners see them; what a User sees depends on the agent's [User Permissions](/core-concepts/agent_access#user-permissions).
| Tab | What it holds |
| - | - |
| **Agent** | What the agent is made of: model, instructions, connectors, skills, knowledge sources, subagents, secrets, abilities. |
| **Triggers** | How it starts on its own: app, webhook, scheduled, and one-time triggers. |
| **Access** | Who can use it and what they can see: Owners, Users, General Access, File Sharing, Task Visibility, User Permissions. |
| **Settings** | Identity, suggested prompts, advanced AI settings, summarization, model fallback, and deletion. |
Sections open in place. Click a section header (for example **Connectors**) and it expands into a full view with an `Agent > Connectors` breadcrumb, so you can drill into a connector's tools and come back without losing your place. Nothing saves until you press **Save** (`Ctrl/Cmd + S`); the arrow next to it undoes unsaved changes.
The model picker and the instructions box sit at the top, with **Edit Own Instructions** ([self-improving instructions](#self-improving-instructions)) right under them. Below that, one section per building block — each shows how many items it has, and **+ Add** to add one.
Lists the triggers on the agent and the five ways to add one. The **Mine / Team** filter switches between the triggers you own and everyone's. **AI Managed** in the header lets the agent set up its own triggers during a task. See [Triggers](#triggers).
Owners, Users, General Access, File Sharing, Task Visibility (team agents), and the eight User Permissions. See [Agent Access](/core-concepts/agent_access).
Identity (icon, name, description), Suggested Prompts, [Advanced AI](#ai-advanced-settings), [Chat Summarization](#ai-advanced-settings), [Model Fallback](#ai-advanced-settings), and **Destructive**, where you delete the agent.
### Guided setup
A brand-new agent asks what it should take care of and sets itself up from your answer, instead of dropping you into the full builder. If you would rather start in the builder every time, turn on **Skip guided agent setup** in [your profile preferences](/core-concepts/profile_settings#preferences).
The rest of this page covers each part, starting with the ones you will reach for most.
***
## Connectors
**Connectors** are the integrations your agent connects to, such as Gmail, Salesforce, Slack, Notion, and 150+ more. The connectors you add determine what the agent can access, so this is the most important part of configuration.
Click **+ Add** on the Connectors header to open the picker. The **All** tab lists every available connector (Gumloop-managed and your own), and the **Custom** tab filters to [custom MCP servers](/nodes/mcp/custom_mcp_servers) you have added. Pick the few your agent actually needs rather than connecting everything.
The **AI Discovery** toggle on the Connectors header is the same setting as [Tool Discovery](#abilities) in Abilities. It lets the agent load tool schemas on demand instead of all at once, which keeps context lean when you connect many connectors.
### Account selection
Click a connector to open its detail view. The **Account** selector controls which connected account the agent uses when calling its tools.
* **Use Personal Default**: each person who runs the agent uses their own default account. This is the default.
* **Use Specific Account**: pin one account for the agent, useful when you have multiple accounts for the same service.
* **Use Team Default** (team agents only): everyone on the team uses the same shared account.
Connect personal or team accounts and choose which account the agent uses.
### Tool Management and approvals
The same detail view has a **Tool Management** panel that controls which of a connector's tools the agent can call, and which ones need your approval first. Tools are grouped into **Read-only tools** and **Write/delete tools**.
Set a preset for the connector, or control each tool individually:
| Preset | Behavior |
| - | - |
| **Always allow** | Every tool runs without asking. |
| **Ask each time** | The agent pauses for approval before any tool call. |
| **Ask for writes/deletes** | Read-only tools run freely; write, delete, and unknown-risk tools pause for approval. |
| **Custom** | Set the mode per tool: allow always, ask each time, or never allow (deny). |
When a tool needs approval, the agent pauses mid-task and shows you an approval card. This is the human-in-the-loop guardrail for sensitive actions.
How approvals work, what the agent shows you, and how to tune when it pauses.
***
## Skills
**Skills** teach your agent how to do specific work *your way*: multi-step processes, templates, and domain knowledge that load only when relevant. You rarely write one by hand. The easiest path is to prompt the agent in a task: get it to do the work well, then say *"turn this into a skill."* When it gets something wrong later, give it feedback and it updates the skill. The **AI Skill Editing** toggle is what lets it create and edit skills on its own.
You can also add a skill yourself with **+ Skill**:
* **Create With AI**: describe the skill and the agent generates it.
* **Upload Files**: turn a document or `.zip` into a skill.
* **Write Skill Instructions**: enter a name and description yourself.
* **Add Existing Skill**: attach a skill you already created.
Create skills, attach them to agents, and build a library that improves over time.
***
## Knowledge sources
Give your agent a searchable memory of what your company knows. In the **Knowledge Sources** section, attach [Brain](/core-concepts/brain) sources (Google Drive, Notion, Slack, GitHub, Confluence, or uploaded files) so the agent answers from your real documents and messages, with citations, instead of guessing.
Click **+ Source** to attach whole sources, or drill into the exact files and folders that matter. When you ask about internal knowledge, the agent searches automatically (shown as *Searching Company Brain*) and can open a full document for more context.
Connect knowledge sources, keep them synced, and give your agents company knowledge to search.
***
## Triggers
Agents can run on their own, without anyone starting a task. The **Triggers** tab lists the triggers on the agent, and **AI Managed** in its header lets the agent create, edit, and manage its own triggers during a task.
There are five ways to add one:
| Option | What it does |
| - | - |
| **App Trigger** | Runs the agent when an event happens in another app (new email, new Slack message, a record changes). |
| **Webhook** | Runs the agent when another service sends it a request. |
| **Scheduled Trigger** | Runs the agent on a recurring schedule, for example every weekday at 9 AM. |
| **One-Time Trigger** | Runs the agent once at a specific time. |
| **Create With AI** | Describe what you want and the agent builds a custom trigger for you. |
Owners manage every trigger on the agent. Users can create and edit **their own** triggers while **Create triggers** is allowed in [User Permissions](/core-concepts/agent_access#user-permissions). The **Mine / Team** filter switches between the two views.
Set up app and scheduled triggers, write prompt templates, and manage active triggers.
***
## Subagents
Subagents let your agent delegate to other agents. Instead of doing everything in one task, it can spin up focused helpers that work in parallel, then collect the results and continue.
There are two ways an agent delegates, both through the **invoke\_agent** tool:
The agent clones itself, keeping the same tools and instructions. Useful for parallel work: spawn several clones, each handling a different subtask. The clone shows up as **"(Me)"** in the list and is enabled by default. Clones cannot clone themselves again (depth limit of 1), and you can scope a clone to a subset of apps.
The agent calls a different, specialized agent by name. Add agents to the **Subagents** list to allow this. Cross-agent chains are not depth-limited, so a coordinator can delegate to a chain of purpose-built agents.
Each subagent runs as its own task with its own context and computer, visible in your task history. A task can have at most **10 concurrent subagents**, including subagents started by other subagents. Subagents also count toward your account's concurrent-agent limit, so available capacity may be lower.
If the agent reaches the subagent limit, it receives "Subagent limit reached." Wait for running subagents to finish, or ask the agent to split the work into fewer batches before starting more.
Behind the scenes, subagents run as queued background tasks with their own time budget (about half the parent's). For batch invocations a shared progress board tracks each one, and the parent can hand specific files to a subagent before it starts. The parent reads each subagent's results when it finishes.
***
## Abilities
**Abilities** are the agent's built-in capabilities. Most are on by default, and each can be toggled from the **Abilities** section.
| Ability | Default | What it does |
| - | - | - |
| **Web Search** | On | Searches the web for current information. Choose the provider (Exa, Parallel, Firecrawl, or the model's native search). |
| **Web Fetch** | On | Reads the content of a specific URL. Choose the provider (Firecrawl, Parallel, Exa, or Gumloop). |
| **Image Generation** | On | Creates images from text prompts. See [Image Generation](#image-generation). |
| **Search Past Conversations** | On | Searches and retrieves earlier tasks for context. The backbone of an agent that learns over time. |
| **Ask Question** | On | Lets the agent pause and ask you a structured, multiple-choice question when it needs a decision. |
| **Tool Discovery** | Auto | Loads tool schemas on demand. In **Auto**, the agent loads tools directly when they are small (roughly 10% of context) and switches to on-demand discovery when they grow larger. |
| **App Rules Creation** | On where available | Lets the agent propose [App Rules](/enterprise-features/app-policies/app-rules) for your review while it works. Enabled by default on plans that include App Rules. |
***
## Image generation
The **Image Generation** ability creates images from text prompts. New agents have it on by default and use **GPT-Image 2.5 Flare** as the default image model.
On an agent, you choose the image **model**. The agent chooses the size, quality, background, format, aspect ratio, and resolution for each request based on the prompt. It is instructed to use the smallest option that satisfies the request.
### Image models
| Picker label | ID | Provider | Per-request options |
| - | - | - | - |
| **GPT-Image-2.5 Flare** | `gpt-image-2.5-flare` | OpenAI | `size`, `quality`, `background`, `output_format` |
| **GPT-Image-2.5 Sunburst** | `gpt-image-2.5-sunburst` | OpenAI | `size`, `quality`, `background`, `output_format` |
| **GPT-Image-2** | `gpt-image-2` | OpenAI | `size`, `quality`, `background`, `output_format` |
| **Gemini 3.1 Flash (Nano Banana 2)** | `gemini-3.1-flash-image-preview` | Google | `aspect_ratio`, `image_size` |
| **Gemini 3 Pro** | `gemini-3-pro-image-preview` | Google | `aspect_ratio`, `image_size` |
| **Gemini 2.5 Flash Image** | `gemini-2.5-flash-image` | Google | `aspect_ratio` |
The **Recommended** image preset tries these models in order: GPT-Image 2.5 Flare, GPT-Image 2.5 Sunburst, Gemini 3.1 Flash, Gemini 3 Pro, GPT-Image 2, and Gemini 2.5 Flash Image. It skips any model your organization restricts.
### Per-request options
The agent selects these options for each image request:
| Option | Values |
| - | - |
| **Size** (OpenAI) | `auto`, `1024x1024`, `1536x1024`, `1024x1536`, `2048x2048`, `2048x1152`, `1152x2048`, `3840x2160`, `2160x3840` |
| **Quality** (GPT-Image 2.5 Flare and Sunburst) | `auto`, `low`, `medium`, `high`, `xhigh`, `max` |
| **Quality** (GPT-Image 2) | `auto`, `low`, `medium`, `high` |
| **Background** (OpenAI) | `auto`, `transparent`, `opaque`. `transparent` requires `png` or `webp` output. |
| **Output format** (OpenAI) | `png`, `jpeg`, `webp` |
| **Aspect ratio** (Gemini 3.1 Flash) | `auto`, `1:1`, `2:3`, `3:2`, `3:4`, `4:3`, `4:5`, `5:4`, `9:16`, `16:9`, `21:9`, `1:4`, `4:1`, `1:8`, `8:1` |
| **Aspect ratio** (Gemini 3 Pro and Gemini 2.5 Flash Image) | `auto`, `1:1`, `2:3`, `3:2`, `3:4`, `4:3`, `4:5`, `5:4`, `9:16`, `16:9`, `21:9` |
| **Image size** (Gemini 3.1 Flash) | `auto`, `512`, `1K`, `2K`, `4K` |
| **Image size** (Gemini 3 Pro) | `auto`, `1K`, `2K`, `4K` |
Agents can also pass reference images to edit them or keep a subject, product, or style consistent. OpenAI image models accept up to 16 reference images. Gemini 3.1 Flash and Gemini 3 Pro accept up to 14, and Gemini 2.5 Flash Image accepts up to 3. OpenAI accepts JPEG, PNG, and WebP; all Gemini image models also accept HEIC and HEIF.
**What it costs:** Image generation is charged from the usage reported by the image model. Cost depends on the model and the options selected, so higher quality or resolution costs more credits. See [Credits](/core-concepts/credits) for how BYOK affects image generation.
Older image model IDs are no longer in the picker. Requests that use a retired ID are remapped to a current model:
| Retired ID | Now uses |
| - | - |
| `dall-e-2`, `dall-e-3`, `gpt-image-1`, `gpt-image-1.5` | `gpt-image-2.5-flare` |
| `gemini-2.5-flash-image-preview` | `gemini-2.5-flash-image` |
***
## The agent's computer
Every agent has a built-in **computer** for running Python and shell commands in a secure, isolated environment. It is always on, so the agent can analyze data, generate files, and run scripts automatically. You do not need to configure anything. It keeps installed packages and workspace files across tasks and ships with 80+ Python packages preinstalled.
What the agent's computer can do, persistence, execution limits, preinstalled packages, and Agent Secrets.
***
## The agent's browser
Agents can also get a real Chromium **browser** to open websites, sign in, fill forms, and download files on sites no connector covers. Turn it on from the **Browser** section of the agent's configuration, watch it live from the chat, and replay every step afterwards.
Turn on the browser, watch it live, answer handoffs, logins, browser profiles, and 1Password.
***
## Measuring and improving an agent
**Performance** in the agent sidebar is where you see how the agent is doing and turn on the two features that improve it: **Evaluations** grade completed tasks against your criteria, and **Reflections** have the agent review past tasks and suggest improvements you can accept or dismiss.
Insights, Evaluations, and Tasks for a single agent.
Define criteria and grade the agent's answers so you catch regressions early.
Let the agent review its own recent tasks and propose improvements.
Agents can also produce rich outputs such as documents, spreadsheets, and interactive charts. See [Agent Artifacts](/core-concepts/agent_artifacts) for how those are generated, shared, and versioned.
***
## Working in a task
### Message input menu
The **+** menu in the message bar gives you quick actions while you work:
* **Add photos & files**: attach files and images to your message.
* **Use skill**: manually point the agent at a specific skill.
* **Mention integration**: reference a connected app directly in your message.
* **Mention secret**: reference a stored secret so the agent can use it without you pasting the value.
* **Incognito**: toggle a private task (covered below).
### Voice input
You can send audio instead of typing. Gumloop transcribes it on the server and sends the text to the agent, so the task flows the same whether you type or talk.
Supported formats include mp3, mp4, m4a, wav, and webm, up to 25 MB. Transcription runs on Gumloop's servers using OpenAI transcription models (Whisper and GPT-4o Transcribe), so your agent only ever receives the text transcript, never the raw audio.
For a spoken, two-way conversation instead of a single dictated message, use [Talk mode](/core-concepts/talk_mode).
### Message queue and steering
You do not have to wait for the agent to finish before sending your next message. While the agent is working, press **Enter** to add a follow-up to the **message queue** without interrupting its current step. You can edit, reorder, or remove a queued message before it is delivered.
To redirect the current run immediately, click **Send Now** on a queued message, or press **⌘ Enter** on Mac / **Ctrl Enter** on Windows or Linux. The agent interrupts its current step and continues with your new input instead of ending the task.
Steering does not undo actions that have already completed, and it does not bypass a pending approval. Use the stop control when you want to end the run rather than redirect it.
### Context usage meter
The circular meter in the bottom-right of the message input shows how much of the agent's usable context window is in use — for models where you picked a [short context window](#context-window), that smaller window is the denominator. Hover it for a breakdown across System, AI Instructions, Abilities, Tools, Skills, Subagents, and Conversation.
If context fills up, reduce the tools or skills attached, switch to a model with a larger context window, or rely on auto summarization to compress older messages.
### Incognito mode
Incognito tasks are **not saved to the database**. They are held in temporary memory and auto-deleted after 24 hours. Toggle **Incognito** from the message input menu before sending.
| Behavior | Standard task | Incognito task |
| - | - | - |
| **Message storage** | Saved permanently | Not saved |
| **Visible in history and search** | Yes | No |
| **Included in data exports** | Yes | No |
| **Files and artifacts** | Stored permanently | Auto-deleted after 24 hours |
| **Used for reflections** | Yes | No, excluded |
Incognito applies to the whole task, including any subagents it spawns. Once it expires, messages and files are permanently gone.
***
## Understanding credit costs
Agents consume credits for AI model usage, tool calls, the compute time they spend working, and an orchestration fee on top of those. Cost depends on the model, message length, task history, and the number of tools available.
The task's running credit total sits next to its name in the header. Select it to open the task's details drawer, which shows the source, creation time, model, and participants, plus a row for each credit type the task was actually charged for: **Chat & Reasoning**, **Tool Calls**, **Compute**, **Orchestration Fee**, and **Evaluation & Self-Improvement**.
Full model pricing, integration costs, and how to track usage.
***
## Settings
The **Settings** tab covers how the agent presents itself and how the model behaves.
| Section | What it controls |
| - | - |
| **Identity** | The agent's icon, name, and description. |
| **Suggested Prompts** | Prompts shown on the agent's home screen so people know what to ask. |
| **Advanced AI** | Max steps, parallel tool calls, max output tokens, and context window. See [AI advanced settings](#ai-advanced-settings). |
| **Chat Summarization** | When and how older messages are compacted as a task gets long. |
| **Model Fallback** | Which model the agent switches to if the primary one is unavailable. |
| **Destructive** | Delete the agent. |
Two things that used to live here have moved: file sharing for generated files is now **File Sharing** on the [Access](/core-concepts/agent_access#file-sharing) tab, and Slack behavior is now under **Channels → Slack**. **Secrets** is a section on the **Agent** tab.
***
## Channels
An agent is not limited to Gumloop. **Channels** in the agent sidebar is the single place you make it reachable somewhere else — every surface is set up from there, one per item in the list.
| Channel | What it gives you |
| - | - |
| **Email Inbox** | The agent gets its own email address. Email it, forward mail to it, or CC it on a thread. |
| **Slack** | Talk to the agent in Slack channels and DMs. |
| **Microsoft Teams** | Talk to the agent in Teams. |
| **Hosted Page** | A public or private chat page you can link or embed. |
| **Developer** | The API and SDK details for calling the agent from your own code. |
| **MCP** | Expose the agent as an MCP server so other AI tools can call it. |
Deploy agents to Slack channels for team-wide access.
Deploy agents to Microsoft Teams channels.
Give your agent an inbox it can read and reply from.
Share a public or private hosted page for your agent.
***
## Finding agents
The **Agents** page lists every agent you can access. Use the tabs to switch views:
| Tab | What it shows |
| - | - |
| **Mine** | Agents you created. |
| **Shared with me** | Agents others shared with you directly or through your organization. |
| **Organization** | All agents visible to your organization. |
Each card shows the agent name, connected apps, creator, and last activity. You can search by name and switch between grid and list views.
***
## Usage stats
A line under the agent's description summarises how it has been used over the **last 31 days**: how many tasks it worked on, and on a team agent, who used it most (**Used by Ana, Ben and 4 others · 62 tasks in the last month**, with avatars for the top users). A personal agent reads **Worked on 62 tasks in the last month**.
Stats respect [Task Visibility](/core-concepts/agent_access#task-visibility). If you can only see your own tasks on an agent, the numbers cover just your usage and no users are named. Subagent runs are excluded, and [incognito](#incognito-mode) tasks never count.
For the full picture — credits over time, credits by model and member, connector and skill usage — open **Performance** in the agent sidebar. See [Agent Performance](/core-concepts/agent_performance). Organization-wide reporting lives in [Insights](/enterprise-features/organization_insights).
***
## Managing tasks
Every task appears under **Recents** in the sidebar. Right-click a task or open its three-dot menu to manage it.
| Action | What it does |
| - | - |
| **Share** | Share the task as a read-only link. |
| **Rename** | Give the task a custom name so you can find it later. |
| **Delete** | Permanently remove the task. |
Rename tasks to keep your sidebar organized. "Q2 Marketing Plan" is easier to find than an auto-generated title.
***
## Guardrails
For agents that take real actions, layer on guardrails so they stay within bounds.
Make the agent pause for approval before write, delete, and other sensitive tool calls.
Set conditions that block or flag specific tool calls at the agent or organization level.
***
## Self-improving instructions
Your agent can update its own instructions while it works. Correct it once ("always check Salesforce first," "keep emails under 100 words") and it edits its instructions so the same mistake does not happen again. Changes take effect on the next step and persist across future tasks. The toggle sits below the system prompt editor and is on by default; there is no version history, so revert by editing the prompt manually.
***
## AI advanced settings
Most agents never need this. Defaults are already set for a good balance of performance, cost, and reliability. Organization admins can set the defaults that new agents inherit. See [Agent Default Settings](/enterprise-features/agent_default_settings).
Open the **Settings** tab, then expand **Advanced AI**, **Chat Summarization**, or **Model Fallback**.
Per-provider model parameters. Switching models preserves your preferences for each provider.
| Setting | What it does |
| - | - |
| **Max Steps** | How many actions the agent can take before responding — each tool use is one step. Default 100, range 1 to 200. |
| **Verbosity** | Length and detail of the response: low, medium, or high. |
| **Parallel Tool Calls** | Whether the model can run tool calls in parallel. |
| **Max Output Tokens** | Upper bound on generated tokens, including visible output and reasoning tokens. Defaults to Auto. |
| **Max Tool Calls** | Cap on total built-in tool calls; further attempts are ignored. Defaults to Auto. |
| **Top P** | Nucleus sampling, 0 to 1. Change this *or* temperature, not both. |
| **Top Logprobs** | How many likely alternatives to return per token, 0 to 20. |
| **Context Window** | How much of the model's context the agent uses before older messages are summarized. Only shown for models that offer two windows. |
| **Thinking** | How hard the model thinks before it answers, and whether it thinks at all. Shown only for models with a thinking control — see [Thinking controls](/core-concepts/ai_models#thinking-controls). |
**Reset Model Parameters to Defaults** restores these values for the current provider.
Some models ship two context windows. When your agent is on one of those models, **Context Window** appears under Advanced AI.
| Choice | What it does |
| - | - |
| **The larger window** (default) | Uses the model's full context. Keeps the most history. |
| **The smaller window** | Summarization starts earlier, which lowers cost and latency. |
| Model | Short | Long |
| - | - | - |
| GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol, Terra, Luna | 272K | 1.05M |
| Claude 5.5 Opus, Claude 5 Opus, Claude 4.8 Opus, Claude 5.5 Sonnet, Claude 5 Sonnet, Claude 4.6 Sonnet, Claude Fable 5.1, Claude Fable 5 | 300K | 1M |
| Grok 4.7, Grok 4.6 | 200K | 500K |
Every other model has a single window and never shows the setting. The choice is saved per provider. It is not sent to the model provider: it changes the context budget Gumloop enforces.
GPT-5.6 and Grok 4.6/4.7 charge higher long-context rates once a request's input tokens cross the short window, and the higher rates apply to the whole request. Claude's 1M-token models bill at one flat rate.
When a task approaches the model's context limit, older messages are compacted into a structured recap while recent messages are kept in full.
Defaults: summarization starts at **80%** of the context window, roughly the most recent 40,000 tokens are protected, and summaries are capped around 30,000 tokens.
Turn on **Override Auto Summarization** to tune **Summarization Trigger (%)**, **Summary Model**, **Protected Context Tokens**, and **Max Summary Tokens**. A lower trigger summarizes sooner and costs less per turn. A higher trigger keeps more raw history.
If the primary model is unavailable or unreachable, the agent retries and then switches to a fallback from a different provider. **Enable Model Fallback** is on by default. Turn on **Override Auto Fallback** to pick fallback models yourself.
Browse the full model catalog by tier, vision models, and bring-your-own-key (BYOK).
***
## Best practices and troubleshooting
Begin with two or three apps and short instructions. Test, watch how the agent behaves, then add tools and rules based on real usage. Avoid launching with 15 tools and a 2,000-word prompt.
When the agent makes a mistake, ask it: "What could I add to your instructions to prevent this?" Then have it update its own instructions or a skill. Review past tasks for patterns.
Be explicit about what the agent should never do without approval: delete records, send emails, make purchases, or modify production data. Back this up with [Tool Management](#tool-management-and-approvals) approvals.
Make tool names more descriptive, add "when to use" guidance in the system prompt, and reduce similar tools that might confuse it.
Check the app's [account selection](#account-selection), authenticate on your [Connectors page](https://www.gumloop.com/settings/profile/connectors?view=connected), and set it as your personal default. For team agents, contact your team admin.
Tune the app's [Tool Management](#tool-management-and-approvals) preset, or list in the system prompt which actions need approval and which can proceed automatically.
Switch to a cheaper model for simple tasks, add your own provider API key (BYOK) to waive model credits entirely, start new tasks instead of long threads, reduce the tool count, and write clearer prompts to cut back-and-forth.
Check that every required app is connected and authenticated, review the task for the specific tool error, and raise **Max Steps** under **Settings → Advanced AI** if the task needs many tool calls in sequence.
***
## Next steps
Run agents automatically on a schedule or in response to events.
Build reusable knowledge packs that teach agents how to do specific work.
Give agents company knowledge to search from your connected sources.
Grade agent responses and catch regressions before they ship.
Let agents review their own work and propose improvements.
Run code securely and manage agent secrets.
Owners, Users, User Permissions, and Task Visibility.