Launching today
Foglamp
Ship AI agents you can actually see
75 followers
Ship AI agents you can actually see
75 followers
The open source observability layer for AI agents built on the Vercel AI SDK. Costs, latency, tokens, distributed traces, evals, and alerts for every generateText / streamText call — in two lines.
Foglamp
Agent observability feels like it will become mandatory. For business agents, the question is not only cost/tokens, but “what did the agent try, why did it decide that, and when did it need human help?” Are you thinking about business-level traces like handoffs, approvals, and failed workflow outcomes?
Foglamp
@rahulbhavsar Great question, and that's exactly the layer we care about. Today every run is a full span tree, so you can already see what the agent tried, the tool calls it made, and where it failed — step by step, down to the token. Multi-step pipelines group under a workflow, conversation turns under a session. The "why did it decide that" part comes from capturing reasoning + the exact prompt/streamed response on every call, plus LLM judges that score things like tool selection and groundedness.
The part you're pointing at — explicit handoffs, approval gates, failed-outcome states as first-class business events — isn't a dedicated view yet. Right now you'd model those as tool calls / workflow steps with metadata. Turning them into proper business-level traces is squarely on the roadmap. Really appreciate the framing.
Thanks, Gustavo. That is a strong answer. The span tree plus prompt, streamed response, tool calls, and judge scores already gets much closer to a real audit trail than the usual “agent activity log.”
The business-level trace layer is exactly what I was thinking about. For AI employees in ops, sales, or support, teams eventually need to review outcomes in their own language: approval requested, handoff triggered, customer risk detected, CRM field changed, escalation failed. Modeling those as workflow steps works early, but turning them into first-class events could make Foglamp much easier for non-technical leaders to trust. Excited to see that roadmap develop.
@gustavofior Visibility is a big missing layer for agents. It’s hard to trust an autonomous workflow if you can’t see what it tried, where it failed, and what changed along the way. Agent observability will probably become a default expectation.
Foglamp
@alpertayfurr Couldn't agree more — that's the whole reason it exists. "What it tried, where it failed, what changed" is literally the trace view. We even fingerprint the model on every call, so you catch silent weight changes when a provider swaps something under you. Thanks for the support 🙏
Congrats on launch! I would love to know how is it different to other observability platform like langfuse/Arize?
Foglamp
@ashishkingdom Thanks! Fair question. The biggest difference is focus: Foglamp is built specifically for the Vercel AI SDK. Instrumentation is two lines — registerTelemetry(foglamp()) — and you get full nested traces, per-agent rollups, cost, and evals with zero manual span wiring.
Two other things set it apart: (1) quality + cost live in one place — evals (code checks and LLM judges) run against real production traffic, not a static test set, and cost is computed per call / agent / customer from live pricing; (2) alerts evaluate every minute, so you find a regression from a dashboard, not a customer. It's also Apache 2.0 and self-hostable with docker compose up. Langfuse/Arize are great general-purpose tools — we're narrower and deeper for teams shipping on the AI SDK.
The 'costs doubled, answers worse, then customers started complaining' sequence is the most relatable AI horror story I've seen this year. Finally something that catches it before the Twitter thread starts. Congrats on the launch!
Foglamp
@laraib Thank you. That sequence is literally the story we built the landing page around: ships clean week 1, costs double and answers get worse by week 3, complaint rolls in week 4. We caught a 10× cost regression on our own stack 3 days after shipping. Catching it before the X thread starts is the entire point 😄 Appreciate it!