Since launching Restless, our most-used feature has been Agent Recovery. It automatically inserts AI-generated recovery steps into error responses, so AI can fix issues more quickly. Using it can cut down on retries or failures significantly, and get agents moving forward in a fraction of the time and tokens.
We're launching Agent Recovery 2.0, which improves the quality of Agent Recovery answers and cuts down on the need for human approvals. We now auto-generate answers to new errors when they first appear, use your codebase to guide the answers, and have tightened how we strip PII and other sensitive data.
Autopilot mode
Previously, when a new error appeared, you'd have to triage it yourself. You'd go in, click "Generate Response", and the AI would generate a reply you could save.
For dozens of errors, this became unwieldy. So, we now can auto-generate recovery steps for each error fingerprint the first time we see it appear, as well as updating stale ones we detect are out of date. This means less work for you, and quicker fixes for each new error.
While we auto-generate answers, humans can still approve them. This will tag the answers as reviewed and approved by a human, so that you can keep organized as to what was generated and what was human-approved.
Codebase-supported answers
Previously, the AI generated the replies from what it had access to: docs, support replies, and the OAS file. While it was able to answer a lot of questions, it fell short on the more esoteric errors: the things that aren't documented anywhere, and the reason for the error isn't clear.
If you connect your GitHub codebase, new errors are now checked against the code to understand what's going wrong. The generated recovery steps are much more helpful, since the agent gets more insight into what's happening behind the scenes.
These errors aren't documented anywhere and existing LLMs have no prior context for them, so without codebase access the recovery steps would be guesses. This gives the recovery steps a much deeper ability to correct usage, which will help agents get back on their way faster.
Security
All answers are pre-generated and no user-submitted content is used in the generation, so there's no prompt injection risk. When a new error is found, we only use the fingerprint and metadata. And just to be completely safe, we use a classifier to detect for any potential prompt injection attempts, and reject anything suspicious.
For auto-generating responses, we use the same protocol we use for any context generation:
- First, we strip all inputs of PII, tokens, or anything that could become an issue.
- Next, when we generate the answer, we make sure to not include any sensitive data.
- We then run an audit with two separate models of the generated answer, to make sure no PII or sensitive data made it through.
- Lastly, we use a classifier to detect if anything generated contains PII or sensitive data, and reject it if it does.
Agent Recovery 2.0 means agents recover faster, with fewer retries, and with answers that are actually grounded in your code. Try it out, and let us know if you have any requests or feedback.
GET STARTED
Build with Restless
The modern API toolkit: docs, an API reference, and AI chat, generated from your OpenAPI spec and updated as it ships.
More from the blog