Gumloop AI Agent Playbook: Connectors, Triggers and Credits

A Gumloop AI agent is a chat agent you describe in plain language and then equip with five things: instructions, a model, connectors to your apps, skills, and Brain knowledge, plus optional subagents and triggers. You can build one in the agent editor or simply ask Gumball, the personal agent every account gets. All of it runs on one credit pool: Pro costs $37 a month for the whole organization's shared pool of 20,000 credits.^[1] The catch is variance. The same agent can spend 2 credits on a quick question and 200 on deep research. So the playbook is mostly about settings. Who can the agent act as, what may it do without asking, and what wakes it up when nobody is watching?

Let Gumball draft the first agent, then take the wheel

Gumball needs no setup. It already uses your own connectors and skills, answers in Slack DMs, and can create or update a custom agent from a chat once you accept an approval card. It cannot delete agents or create app triggers, so those steps stay with you in the editor. If you would rather start from a working example, Gumloop's use-case gallery loads a preconfigured agent with its instructions in place and its apps ready to connect; our guide to Gumloop's ready-made agents walks through the five categories.

Two defaults deserve a look on day one. First, the agent can edit its own instructions, the toggle is on, and there is no version history, so keep a copy of prompts you care about. Changes to its connectors, skills and triggers each need your approval. Second, new agents reply in Classic mode, one final answer per task; Conversational mode sends several messages while it works. Gumloop's older step-by-step builder, Workflows (once called Flows), still appears on the pricing page marked Legacy; the current docs and product are built around agents.

Connectors decide what the agent can touch

Connectors are the most important part of the configuration. You add each one on the agent itself and sign in; connecting an app in Settings alone does not attach it. Gumloop lists 300+ built-in connectors plus any MCP server, and the agent does not load every tool up front. It searches its connectors and loads only what a step needs, so an extra connector adds little cost per step. Our Gumloop MCP guide covers custom, hosted and proxied servers.

Account mode matters as much as the connector. The default is personal: each person who runs the agent acts through their own account. Team agents can use a team default, an agent can pin several named accounts, and agent-owned credentials need Enterprise and a custom role. With pinned accounts the agent sees each one by name, uses the default unless a request points elsewhere, and can read from one account and write to another in the same task. Pick the mode before you share the agent.

Then set approvals. Every connector starts on Always allow. The other presets are Ask each time, Ask for writes/deletes, Never allow and Custom per tool. Gumloop sorts each tool into read-only or write when you add it, so the writes preset works on custom connectors too. Rejecting is non-destructive. For sandbox code, the approval card shows the exact code. Conditional App Rules are an Enterprise policy.

Model, skills, subagents and Brain

Choosing the model

Every agent picks Auto, Balanced, Smartest or a specific model from 35+ models by Anthropic, OpenAI, Google, Meta, xAI, Moonshot, DeepSeek, MiniMax and Z.ai. Auto is Gumloop's router. It picks a lower-cost model that can handle each message, can step up mid-run and shows which model ran. You pay the tokens of that model; the routing decision itself is free. Writes to real apps, subagent turns and images run on a mid-tier model or above. Each trigger can carry its own model, and a heavier pinned one costs more per run.

Skills versus subagents

A skill is a folder with a SKILL.md file, plus optional scripts and references, that the agent loads only when relevant. Until then it sees names and descriptions, roughly 50 to 100 tokens per attached skill. Keep SKILL.md under 500 lines. A skill runs inside the current chat. A subagent is a separate agent with its own tools and context, built for parallel or specialist work: up to 10 at once per task, each working for up to one hour. One that exhausts its step or credit budget pauses for your approval instead of failing.

Brain as memory

A new conversation starts empty. Brain, on Pro and Enterprise, indexes company knowledge from Drive, Notion, Slack, Confluence^[p1], Salesforce and other sources, searched with citations. It has no OCR and skips private Slack channels. New sources get a free estimate and wait for approval before indexing spends credits. After that, indexing charges only for new or changed documents, and each search is billed per query. Sources can sit at personal, team or organization scope, so a sales agent need not see finance files.

Triggers are where agents spend while you sleep

An agent without a trigger waits for you. With one, it works on its own, and that is where unplanned spend appears. Scheduled triggers run as often as every minute. App triggers for Gmail, Slack, Salesforce, HubSpot, Zendesk and others either poll about every 60 seconds or fire within seconds in real time. Triggered runs bill model and tool usage exactly like a chat.

AI-created custom triggers are the expensive kind to leave running. They combine several services and computed conditions, poll every 5 minutes by default, which is also the shortest interval (you can slow one down to once a week), and bill every poll whether or not anything fires.^[2] Sandbox polling costs at least 1 credit a poll, and any decision-model call adds more. Left at 5 minutes, that is 288 polls a day, so at least 288 credits a day and about 8,640 credits in 30 days, by our arithmetic. That is roughly 43% of Pro's monthly pool before the agent does any real work.

Three more rules keep triggers predictable. They run with the creator's credentials, though team agents can use shared team credentials instead. A trigger switches itself off after three consecutive failures and notifies the owner. If the previous run is still going, the next one is skipped.

Credits shape triggers too. When they run out, scheduled and app trigger runs fail until you top up or raise the cap, while AI-created triggers pause polling and resume when credits return. Pro credits also reset monthly without rollover (Enterprise credits roll over).

Grade the work, then decide who applies the fixes

Evaluations grade finished tasks against criteria the owner writes and return Pass, Flag or Notify. A failed Notify criterion fires Slack and email alerts; Flag only marks the chat. Each agent allows 30 criteria, 50 tags and 40 data points, and incognito chats, an Enterprise mode, are never graded. Watch your balance here. When credits are short, an evaluation is skipped silently, with no error.

Reflections go further. On a schedule, the agent reviews its recent tasks and configuration and proposes a few concrete fixes with evidence: a new skill, an edited instruction, a trigger change. New agents on Pro and above start with Reflections on, and Auto-Apply Suggestions is the default.^[3] Auto-apply is not limited to low-risk changes. It can change triggers immediately and remove skills. For an agent that writes to a CRM or emails people, switch to Review Queue, which holds every suggestion until you accept it. Unanswered suggestions expire after 30 days, and you cannot run a review on demand.

Credit habits that hold up across many runs

Gumloop's docs name four levers: cheaper models, your own provider key, shorter task threads and fewer tools. Starting a new task instead of extending a long thread matters because every model step is billed, thinking tokens included, and long conversations carry more context into each step.

Your own model key is the biggest lever, with a twist. Model calls the key covers cost no credits, but tool calls and compute stay billed and the orchestration fee rises from 8% to 16% of the full list cost. In the docs example, a 459-credit run (about $2.30) drops to 93 credits (about $0.47), by our arithmetic. So a key saves credits on token-heavy runs but can add credits on runs that are mostly tool calls or browser and sandbox time. Compute bills at 5 credits per minute of active processing, so a long browser or sandbox session adds up even on a cheap model; our web scraping guide covers the agent browser.

Two outputs have their own billing. Files an agent creates become artifacts, and static files do not keep billing. Live artifacts, HTML pages whose scripts run on every open, charge the viewer's credits. The Performance page shows credits per task, model, trigger, member and connector over the last 31 days. Check it after the first week.

When a Gumloop agent fits the job

A Gumloop agent fits work that crosses several company apps, benefits from a person approving the risky step, and can run on one shared credit pool. Build it with Gumball or a gallery agent, attach only the connectors it needs, set writes to ask, and choose Review Queue for Reflections on anything customer-facing. Treat AI-created triggers as a standing cost rather than a free watcher. The agent that stays cheap is the one with a narrow job, short threads and a trigger you have priced in advance.