Glen (YC Summer 2026): teardown vs XMarts

In one line: Glen records what a team and its coding agents do (agent sessions, Slack, GitHub, tickets, calls, docs), then feeds the relevant past work into the next agent's prompt and turns repeated work into shared skills [site] [yc]. Threat: 2, because it is the closest rival to W8 (skills made from real work) and it is live with named startup customers, but it is one person, brand new, and stops at memory [yc] [site].

1. Fact file

2. Their proposition, reverse engineered

3. Positioning

• Category they claim: "Institutional learning layer" for agents; also "company brain" and "multiplayer AI" [site] [yc].
• Category a YC partner would put them in: agent memory and context for engineering teams, with a skills feature on top (inference). Same family as the company-brain bets the brief lists.
• What they want to be known for: the shared memory every agent reads from and writes to, so nothing a team learns is lost.
• Where they sit on the map: L1/L2/L3 = 3/1/1. Buyer: small startups (inference), developer buyer, business-side users in Slack.

4. Three layers vs ours

5. Same and different

6. Wedge overlap

7. Whitespace (where they stop)

• Layer gap: they remember and steer; they do not run a team of agents or check results [site]. To own it we need one customer's agent team run for 4+ weeks with approvals and checks, with a before/after number.
• Buyer gap: only software startups show up, and the site names no UK or Europe customer [site]. To own it we need a non-software buyer (services firm or SME) saying they have the pain, with one named person on file.
• Number gap: their number is tokens and time on one repo, run by them [blog]. No money or output-per-person result. To own it we need a number a buyer would sign, such as shipped work per engineer per week on a customer's repo.
• Price gap and speed gap: not real here. Price is not public (unknown), and they claim setup in under 5 minutes [site]. Do not claim either.

8. For the YC application

• Proven today: Marton's logs show his system already runs agents that do and check work: about 5,600 jobs, each checked by a different AI, with about 160 failure records that each added an automatic check. Glen's public result is a vendor-run replay of 100 PRs on one repo [blog].
• Draft answer to "how are you different from Glen?": DRAFT. "Glen gives agents a memory of what the team did. Our system runs the agents, checks every result with a different AI, and turns each failure into a rule, so the next run is better. We have run that on our own business every day." Founders say it their own way.
• Future claim (needs proof): "Skills that improve because the work was accepted or rejected." Needs one customer team's skills and a log of rule changes tied to accepted or rejected runs over 4+ weeks, with rework falling.
• What we understand that they don't: a team learns more from checked results and failures than from recalling past sessions. needs proof on a customer; Marton's logs show it on his own work (offer feedback lines fell from 44 to 18 as rules were added).
• Copyability (M3): easy. They already read GitHub PRs and tickets [site], so accept and reject signals sit in their data (inference). Adding "learn from accept or reject" is a feature for them, it fits their buyer, and they built a replay test with rollback in two months [blog]. Not hard, so M3 can not score above 3 against Glen on W8 alone.
• Why now (M4): None found in this teardown. Do not use "agents can now do the work": that is general AI progress.
• Do not claim: "we are the only ones who make skills from real work" (Glen does). "Shared company brain" as our headline (Glen and others own the words). Any outside customer result (none yet).

9. Ideas worth taking

1. Prove value with a replay test: roll the memory back, redo real past work with and without the system, report the means over 5 runs [blog]. We can run the same shape on our own repos for a number.
2. "Ask why this line exists" by questioning the agent that wrote it [site]. A cheap way to make a record useful to a buyer.
3. A private mode that keeps a session out of the store [site]. It answers the client-data worry before a buyer asks.

10. Unknowns

• Price and contract size. Ask a design partner, or watch for a pricing page after the waitlist.
• Funding after YC. Check press after Demo Day.
• Real scope of capture: the site says it reads only public Slack channels and shared transcripts in one place, and captures broader sources in another [site]. Ask for the security walkthrough.
• Whether skills are updated or retired from use, or only created. Look at the product docs when public, or watch for a changelog.
• Any UK or Europe customer. Check the customer page and the next blog posts.
• Whether the Composio use is a paid contract or a friend of the founder (inference: he worked there). Look for a Composio case study.

Sources

• tryglen.com [site]
• Glen Vision [site]
• YC company page [yc] [jobs]
• Replaying 100 PRs with and without Glen, 4 Aug 2026 [blog]
• Glen blog index [blog]
• Composio raises $29M for AI learning, Jul 2025 [press] (context on the founder's former employer; not about Glen)