Documentation Index
Fetch the complete documentation index at: https://university.gumloop.com/llms.txt
Use this file to discover all available pages before exploring further.
With Reflections, Gumloop agents can automatically improve themselves, simply based on your conversations. When enabled, in every conversation agents will identify things that went well, things that went poorly, and periodically reflect on those things and suggest improvements. That could be an update to its instructions, a new or updated skill, or an app that should be added to the configuration of the agent. All of this happening in the background.
So let’s go ahead and look at an example. I’ve got this agent from the marketing team that qualifies leads, runs reports, and more. Under Reflections, I’ve got it configured to confirm any changes it wants to make, any reflections that it’s come to, and reflect once every day. I’m not trying to optimize anything specific like credits or speed, so I’m keeping the custom instructions blank, and probably you should too. And I want notifications when new reflections come in via email and Slack.
So let’s go ahead and check back tomorrow and see what it thought of, well, where it can improve.
All right, so we’re back, and let’s see what the agent thought it can improve. It’s found two updates to its instructions that would improve itself. First is to add a summary of the qualification document that it’s reading in every single chat directly to its instructions, and another to skip Apollo tool calls where it’s seeing errors.
Now, when I approve, it’s actually going to send itself a prompt to implement the change. Yes, the agent reads its own summaries, comes up with a prompt to improve itself that it then sends to itself. It’s a new world out there. And now I can see that it’s gone ahead and updated its own instructions based on what it’s reflected on. My agent has improved itself. I know, it’s crazy.
Two final things here. First, Reflections do cost credits. However, we’re working hard on minimizing the costs of the actual Reflections and making sure that the payoff in accuracy and efficiency is worth that investment, that it’s a net positive for you and for your agents. And second, I’ll end by asking for your feedback. There’s a lot of nuance here. So please tell us if we found the right balance. Are the Reflections useful? Are they coming too often? Are they not useful at all? Please let us know.
And now that you’ve reflected on Reflections, it’s time for your agents to do the same. So go ahead and turn it on and let us know how it goes.
Reflections
Reflections let your agents review their own conversations and suggest improvements to themselves, from updated instructions to new skills.
Agents get better with use, but only if someone is paying attention. Reflections automate that process. When enabled, your agent reviews every conversation, notes what worked and what didn’t, and periodically suggests changes to its own configuration.
How it works
After each conversation, the agent logs observations in the background. On a schedule you set (daily, for example), it aggregates those observations and proposes concrete improvements. These could be updates to its instructions, new or revised skills, or apps it should add to its toolset.
You can choose whether the agent applies changes automatically or waits for your approval. When you approve a reflection, the agent writes a prompt for itself and executes the change. It literally rewrites its own instructions.
Setting it up
Open the Reflections panel on any agent. Pick how often you want reflections (once a day is a good starting point), decide whether changes need your confirmation, and optionally add custom instructions if you want to steer what the agent optimizes for. For most agents, leaving custom instructions blank works fine.
You can also turn on notifications via email or Slack so you know when new reflections arrive.
What agents actually suggest
The improvements are specific and practical. An agent might notice it keeps pulling data from a tool that returns errors and suggest skipping those calls. Or it might realize a document it references every chat should be summarized directly in its instructions to save time. These are the kinds of optimizations that would take you weeks to notice manually.
What to remember
Reflections cost credits, but the goal is a net positive: better accuracy, fewer wasted tool calls, and agents that get sharper over time without you having to babysit them. Start with daily reflections, require approval so you can see what the agent proposes, and adjust from there.
