Respan (YC Winter 2024): teardown vs XMarts

In one line: Respan sells AI engineers a gateway, tracing, evals and alerts to watch and fix the AI agents they ship, with a free tier and a $199 a month team plan [site] [yc]. Threat: 3, because it is the clearest W2 rival: same kind of buyer, same "is my agent working?" problem, funded, live, with named customers, and it already traces Claude Code, Codex CLI and Cursor runs [press] [site] [docs].

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: "first proactive AI observability platform" and "control plane" for agent quality [blog].
• Category a YC partner would put them in: LLM observability and gateway (Langfuse, Braintrust, LangSmith, Portkey space) (inference).
• What they want to be known for: observability that does not stop at a dashboard but turns what it sees into a fix.
• Where they sit on the map: layers L1/L2/L3 = 0/1/1, buyer = sme, developer.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: they check; they do not do the customer's work, and they hold no memory of the customer's company [site]. To own it we need a customer's agents to do real work, with the check and the memory in one loop, for 4+ weeks.
• Number gap: their numbers count volume, speed and uptime, not finished work (logs, calls, "90% less debugging time") [site] [blog]. To own it we need one number a buyer signs: work accepted, landed or paid, per week, with cost per unit.
• Buyer gap: not found as a gap. They sell to the same small tech teams, at a low price [site]. Do not claim it.
• Price gap: not found. A free tier and $199 a month is cheaper than anything we would charge [site]. Do not claim it.
• Speed gap: not found. One CLI command and "traces within minutes" [press] [docs]. Do not claim it.

8. For the YC application

• Proven today: Marton's logs show a fleet that has run 5,641 agent jobs, with 864 merged. About 160 incident records each name a failure class and an automatic check. One record shows a single launch that authorised 156,000 agent turns and used 39% of a weekly budget in 14 hours. Respan's public pages count calls and logs, and no page we found counts work landed [site] [blog].
• Draft answer to "how are you different from Respan?": DRAFT. "Respan shows you what your agents did, call by call. We track whether the work landed, and we turn each failure into a check that stops it coming back. We learned this by running our own agent fleet for five months." Founders say it their own way.
• Future claim (needs proof): "We show a small team how much agent work was accepted each week and what it cost per piece." Needs one customer's agent fleet watched for 4+ weeks, with a before/after number, and a build that does not depend on our Macs (inventory weakness).
• What we understand that they don't: the failure that matters is not a bad answer but a job that quietly never lands, and the fix is a rule, not a dashboard. proven on our own fleet; needs proof on a customer.
• Copyability (M3): medium. The tracing and alerting parts are easy for them: they already ship coding-agent tracing, error incidents, monitors and an MCP with 40+ tools [docs] [blog], and their buyer uses these agents (inference). The part that is harder is tying traces to landed work and the named failure classes, which needs the merge side and months of operating history (inference). Not hard, so M3 can not score above 3 against Respan.
• Why now (M4): None found in this teardown. Respan's own dated moves (rebrand 20 Feb 2026, "proactive" launch 18 Mar 2026) show the market is moving to closed-loop observability [blog] [press], but that helps Respan as much as us.
• Do not claim: "no one watches coding agents" (Respan traces four of them [docs]). "Self-driving observability is hype" (we found no autonomous fix claim from them, only suggestions; we have not tested the product). "Cheaper" or "faster to set up" (their price and setup are very low). Any customer count or revenue of our own.

9. Ideas worth taking

1. A "Copy fix prompt" button: hand a failure plus its evidence to a coding agent in one click [blog].
2. Group many failures into one incident and label whose fault it is (provider, your requests, platform) [blog]. Easy to act on.
3. Rank-style content: "best alternative to X" comparison pages that catch buyers who are already searching [blog].

10. Unknowns

• True team size (YC says 23, careers page says about 10). Look at the YC page and their careers page again in a week.
• Revenue and growth. They claim 8x in 2025 with no base [press]. Ask investors or look for a Gradient post.
• Whether their coding-agent tracing shows stuck, retried or never-landed jobs, or only calls. Read respan.ai/docs/integrations/claude-code; ideally run the free tier on one of our own jobs.
• Whether the "evaluation agent" applies prompt changes by itself or only suggests. Docs and launch posts do not say.
• Sales motion. The Founding Technical GTM ad body sits behind a YC login.
• UK or European customers. None seen.

Sources

• respan.ai home [site]
• Respan pricing [site]
• Respan customers [site]
• Respan careers [site]
• Respan docs [docs]
• Respan docs index (llms.txt), coding-agent integrations [docs]
• YC company page [yc]
• YC jobs page [jobs]
• Announcing our new brand: Respan, 20 Feb 2026 [blog]
• Respan Agent, CLI and MCP (Launch Week day 3), 22 Apr 2026 [blog]
• Error tracking in Respan, 13 Oct 2026 [blog]
• Introducing Span-02, 15 Oct 2026 [blog]
• How Octolane scaled to 4.4M agent requests, 30 Sep 2026 [blog]
• Respan blog index [blog]
• SiliconANGLE: Respan raises $5M, 18 Mar 2026 [press]
• VCA Online press release, 18 Mar 2026 [press]
• CB Insights, Keywords AI financials (via search result, not opened) [press]