A SaaS tool that lets solo devs and small AI teams record, tag, replay, and diff agent trajectories to catch prompt-regression bugs without re-running expensive benchmarks.

Customer: Indie AI developer or solo founder who has shipped at least one LLM-powered agent to production (e.g. a coding assistant, research bot, or support agent) and is now iterating on prompts/tools weekly — and keeps accidentally breaking behavior they fixed two weeks ago.

Problem: Every prompt tweak is a gamble. There’s no cheap, fast way to know if changing a system prompt made tool selection worse across the 40 edge-case trajectories you’ve hand-curated over the past three months. Re-running a full benchmark suite costs $20–$80 in API calls and 2 hours of wall time. So most solo devs just… ship and hope.

Pricing: saas-mrr — $800 MRR in 4 months (16 paying users at $49/mo)

Why now

ADRA-Bank and SWE-rebench V2 have just mainstreamed the idea of trajectory-level evaluation — but those are research benchmarks, not developer tools. The gap between ‘benchmark paper’ and ‘usable devtool’ is wide and largely unoccupied at the indie scale. Developers are also running agents on cheaper models (GPT-4o-mini, Haiku) where regression is more frequent, making cheap replay checks genuinely valuable.

Go-to-market

1. Post a working open-source CLI on GitHub that records and replays a single agent trajectory locally (no account needed) — target the ‘show HN’ crowd and /r/LocalLLaMA. This builds credibility and seeds organic discovery.
2. Find 5 indie devs in public AI-builder communities (Latent Space Discord, AI Engineer Slack, Buildspace alumni) who are actively maintaining a production agent. Offer free white-glove onboarding in exchange for a 20-min recorded feedback call and a testimonial.
3. Write one very specific blog post: ‘How I caught a tool-selection regression before it hit production’ — use your own dogfooded agent as the example. Submit to Hacker News, The Batch, and TLDR AI newsletter.
4. Gate the trajectory-diff UI and team sharing behind a $49/mo Pro plan. Free tier = local CLI + 50 stored trajectories. This creates a natural upgrade trigger the moment someone wants to share a regression report with a contractor or second founder.

Moat (or lack thereof)

No real moat. The core ideas are reproducible by any competent Python dev in a weekend. The defensibility is purely in distribution (being first in devtool mindshare for this specific workflow) and in accumulated trajectory datasets that users won’t want to migrate. Expect a larger player (LangSmith, Braintrust, Weights & Biases) to add a similar feature within 12–18 months — the window is narrow but real.