Stacktrace (YC Fall 2026): teardown vs XMarts

In one line: Stacktrace is a control layer for coding agents. It watches what each agent did, flags risky or broken runs, sets one rulebook for the team, and alerts in Slack [site] [yc]. Threat: 2, because it is the closest rival to W2 and has strong founders, but it is three people, brand new, names no customer and checks only (it does not do or measure the work) [site] [yc].

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: "control layer for every agent" [yc]. Control plane for coding agents [site].
• Category a YC partner would put them in: agent security and monitoring (DevSecOps), next to Wato, Salus-style behaviour checkers and Nessie (inference from the brief's list and their YC tags).
• What they want to be known for: the control point for every agent a company runs.
• Where they sit on the map: layers L1/L2/L3 = 0 / 2 / 1, buyer = unknown size, 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 or measure the work. A finding says "this was risky", never "this landed" or "this cost £X to produce" [site]. To own it we need a decoupled W2 tool, run by 3 outside teams, that links agent runs to merged work and spend.
• Number gap: no outcome number in their proof, and cost rules are "coming soon" [site]. To own it we need one number a head of engineering would sign: for example silent failures caught per week, or cost per merged change, over 4 weeks on a real team.
• Not true, so left out: price gap (the Team tier is free) [site]; speed gap (install is one pasted prompt) [docs]. Buyer gap is weak: they start at one developer and reach up to Enterprise [site]. An engine gap exists today (Claude Code only live), but it is a roadmap item for them, not whitespace (inference).

8. For the YC application

• Proven today: Marton's logs show his system tracks what is stuck, what silently stopped and what it costs across three AI makers (5,641 jobs, Codex 4,756, Claude 839, Gemini 46), with about 160 failure records that each added an automatic check. Stacktrace's public pages describe risk rules on Claude Code sessions, ten live, and no outcome or cost measure yet [site].
• Draft answer to "how are you different from Stacktrace?": DRAFT. "Stacktrace watches what agents do and blocks the risky ones. We track whether the work actually landed and what it cost, because we run a fleet of agents every day and that was our pain." Founders say it their own way.
• Future claim (needs proof): "Teams running agent loops see, in one place, which runs produced merged work and what each cost." Needs 3 outside teams on a decoupled W2 for 4 weeks with a before/after number, and the tool running off Marton's Macs.
• What we understand that they don't: the buyer's daily question is "is my fleet producing?", not only "is it safe?" needs proof with an outside buyer; Marton's logs show the failure classes come from his own five months of running a fleet.
• Copyability (M3): easy. Reading session files is what they already do [docs]. Cost rules are already announced [site]. Linking runs to merged work is a small extra step, and the CEO ran engineering at a CI and delivery company (Harness) (inference). Wanting to is also likely, since it fits their buyer. M3 can not score above 3 against Stacktrace.
• Why now (M4): None found. Do not use "agents are everywhere now": that is general AI progress.
• Do not claim: that Stacktrace ignores cost or stuck runs (they have retry-storm rules and cost "coming soon") [site]. That W2 is a shipped product (it is Mac-bound and unused by anyone else). That our edge is hard to copy. That we can out-build three infrastructure veterans on monitoring itself.

M2 (what YC funds now) is NOT judged here. It comes from the themes of many companies in Pass 3.

9. Ideas worth taking

1. A browser sandbox where you drop in a session file and nothing leaves the tab. A zero-install, zero-trust demo [site].
2. "Rules ship only once their findings get acted on" [site]. Same spirit as our ratchets: a check earns its place by being used.
3. Install by pasting one prompt into the agent itself [docs]. Fast first value.

10. Unknowns

• Any customer, revenue or install count. Counters read zero. Look at the GitHub plugin repo and their X account for first users after launch.
• Funding terms beyond YC. Look for a seed announcement after Demo Day.
• When cost and steering ship, and when Codex goes from preview to live. Watch docs.stacktrace.ai "Coming soon" page.
• Whether OpenACA was the company's earlier name or a separate project. The YC data says former name; their About page is silent.
• UK or EU presence. Nothing on the site.

Sources

• stacktrace.ai home [site]
• Fleet [site]
• Personal Agent Gateway [site]
• Pricing [site]
• About [site]
• Sandbox [site]
• Demo booking [site]
• Docs [docs]
• YC company page [yc]
• GitHub org stacktrace-ai: one repo, stacktrace-plugin, 0 stars, updated 19 Oct 2026 [docs]
• AI Indigo review of OpenACA, 10 Jul 2026. It does not name Stacktrace or the authors [blog]
• OpenACA on PyPI: the page did not load; only a search summary was seen. Treat the link between OpenACA and Stacktrace as UNVERIFIED except for the YC "former name" field and the docs' Open source list [docs]
• A news search for launch coverage or funding returned nothing about Stacktrace [press]