Which advantages over its rivals hold up, and why are they hard to copy?
Mirrors holds the delivery advantage over LangSmith and Braintrust because a correction reaches only the conversations its trigger matches, whereas a suggested fix from those rivals still must reach the agent, usually as a prompt or code change.
Developers write a correction, select the situations where it applies, and preview its scope against past conversations. LangSmith and Braintrust can now identify recurring problems and suggest fixes.
Mirrors connects through an SDK and works with any agent framework. The harness retains control of the loop, model provider, retries, streaming, and cancellation.
What could change its growth trajectory, and what would confirm it?
Mirrors' per-correction outcome comparison could raise its growth by giving a team evidence to expand from the Hobby plan—offering 1,000 free monthly turns shared among all the customer's agents—into the paid Growth plan, which charges one monthly base fee, together with the turns its agents use. When a situation matches, Mirrors delivers guidance and compares outcomes with held-out conversations using the outcomes the team tracks.
Langfuse includes 100k units each month and bills $8 per 100k units beyond that, with discounted rates at higher volume. Customers can begin on Langfuse's Hobby tier at no cost without a credit card.
Which risks could break the thesis, and what would warn of them?
Silent non-delivery could undermine Mirrors' paid-guidance thesis, because context retrieval is optional and past 750 ms the system fails open, leaving the agent to run without the rules. Context retrieval returns optional text within a bounded wait, and the application code assembles the prompt.
Mirrors treats receiving context in a turn as separate from capture delivery, the actual prompt the harness sends, added turn latency, and task outcomes. Mirrors recommends starting in shadow mode to verify capture and inspect failures. With active delivery enabled, eligible guidance can reach later turns without another application deployment.
What does the product do for its buyer, and how is it delivered?
Mirrors corrects recurring agent mistakes without growing the system prompt for every conversation: before starting each user turn, the harness sends Mirrors the original conversation and requests relevant guidance, which the system returns to the existing harness.
The harness retains control of the loop and the model provider, along with retries, streaming, and cancellation. Each plan offers unlimited agents without agent fees. Mirrors runs alongside a team's tracing and evaluation tools.