XMarts room

Winter 2023

Backdrop

Building your AI coworkers for any team, workflow, or industry.

Visit website

What they do

platform for AI coworkers that share company memory and hand work between people

How close to us

4, very close

AI coworkers with approvals and shared company memory; agents pass work to each other, matching W9; buyer size not stated.

Wedges

  • W1 Trusted 24/7 coding for small startup teams direct
  • W6 A tool for product managers direct
  • W2 Seeing whether your AI agents actually produce adjacent
  • W3 Company memory from every call, for consultancies adjacent
  • W8 The other 90% of an AI skill adjacent
  • W9 Ask my agent before you book my time adjacent
L1 Company brain (knows the organisation) 2, clear
L2 Agent team (runs agents, tools, approvals) 3, strong
L3 Does and learns (finishes, checks, measures, updates the brain) 1, light

Who buys

Firm size
sme
Buyer type
business
Role
team or department head, founder of a small firm (inference)
UK
not known
Their own description (from YC)

Backdrop is a platform for building AI coworkers around the work you actually need done.

Start with an existing coworker or create your own for almost anything. Research a market, prepare client updates, run outreach, create campaigns, analyze customer feedback, build dashboards, monitor orders, coordinate projects, test products, handle recurring admin, or build a workflow completely specific to your business.

Connect the systems where the work already happens, including Slack, Gmail, Google Drive, HubSpot, Shopify, Jira, Asana, Linear, GitHub, Notion, and more. You control what each coworker can access, what it can do, and which actions need approval.

Backdrop works just as well when one person wants to hand off a job to AI. But it is also built for something today’s 1:1 AI doesn’t handle well: working together.

Employees, AI coworkers, clients, and partners can collaborate in shared conversations and workflows. One agent can bring in another. Work can move between people and agents without everyone starting a new chat and rebuilding the context from scratch.

Backdrop builds shared company memory from the work happening across your company. Decisions, conversations, feedback, documents, project history, and activity across connected tools can become context that people and AI coworkers can use later. A coworker can understand what happened before, another agent can build on what was already learned, and a teammate can understand why something was done without digging through old threads or systems.

The result is AI that works for an individual when that’s all you need, but can also become part of how the whole company works.

Backdrop can support businesses ranging from agencies and e-commerce companies to manufacturers, service businesses, SaaS teams, and more, because you build the coworkers around your work, not a fixed template.

Full teardown

Backdrop (YC Winter 2023): teardown vs XMarts

In one line: Backdrop sells "AI coworkers" that small teams build, connect to their tools (Slack, Gmail, HubSpot, Linear, GitHub) and control with approvals. The coworkers share one company memory and hand work to each other [site] [yc]. Threat: 2, because it is the same buyer and the same idea of an agent team with shared memory, but it is two people, has no named customers, and stops at a human approval and an audit log, not a measured result [yc] [site] [press]. Expect a YC partner to name it for W1 and W6 because it ships a PM agent and an engineer agent [blog].

1. Fact file

Item Finding Source
What they say Home page: "Build the AI team your work needs." YC line: "Building your AI coworkers for any team, workflow, or industry." [site] [yc]
Founded, team, funding YC W23, Seattle. YC shows a team of 2. Funding not public: one third-party database snippet says a $500k pre-seed in Apr 2023 (not opened, UNVERIFIED). [yc]
Founders Akanksha Singh (CS, Univ. of Michigan 2018; worked at Microsoft and Uber; Forbes 30 Under 30 2024, Hollywood and Entertainment) and Caitlin Short (CS and Econ, Michigan 2017; worked at Honey, later bought by PayPal; grew up around film sets). They met at Michigan. What YC likely saw in 2023: two engineers with a real insider view of a messy back office (inference). [press]
Customers None named. The site shows example coworkers (Dinesh, Alex, Sam) and sample dashboard figures that read as mockups. Product Hunt has a Customers tab with nothing in it. One Product Hunt reviewer says they "used Backdrop to build" another product (one review, 5.0). [site] [blog]
Price Public, self-serve. They list four plans, priced per coworker per month (currency not shown): Starter free, Base 100, Pro 250, Pro Plus 550. Credits per plan: 5,000, 30,000, 80,000, 180,000. Top-ups sold. [site]
Path Twig (film, TV and music production expenses, YC demo day Apr 2023) → Backdrop (same product, receipts and production accounting, Mar 2024) → YC tagline "Hire your first AI product team" → Alex (AI product manager) and Sam (AI engineer) launched on Product Hunt 20 Jul 2026 → general "build any coworker" relaunch 15 Oct 2026. "Sphere" is listed by YC as an old name; no date found. [press] [yc] [blog]

2. Their proposition, reverse engineered

Part What they do
Who buys Businesses "from a single workflow to a whole team": e-commerce, manufacturing, agencies, SaaS, service firms [site]. Pricing and a free tier point to small teams (inference).
The pain AI lives in separate chats, people explain the same context again and again, and building useful agents takes setup and upkeep (founder's launch note, paraphrased) [blog].
The promise Build a team of AI coworkers for the work you need done, with shared context and control over what each can do [site] [blog].
How it works Pick a ready coworker or build your own. Connect tools (they claim 1,000+ on the site, 1,500+ on Product Hunt). Add skills, or teach a workflow by showing a recording. Set permissions and which actions need approval [blog] [site]. One agent can hand a task to another; it lands on their queue and the handoff stays on the task [blog].
The output Finished work inside the buyer's own tools (tickets, pull requests, reports, messages, dashboards). A full record of what an agent did and why [blog].
Proof style None in numbers. The site shows sample figures (for example 186 Shopify orders tracked) that look like illustrations, not customer results [site].
Who sells it Self-serve sign-up, plus a "book a demo" link [site]. No open jobs on YC, so no sales hire to read [jobs].
Price model Per coworker per month, plus credits that run out and can be topped up [site].
The story Enemy: 1:1 AI chats that each start from zero. Belief: AI should be part of how the whole company works, with shared memory [yc].

3. Positioning

  • Category they claim: "AI coworkers" / the AI team your work needs [site].
  • Category a YC partner would put them in: horizontal agent-builder platform for small and mid-size teams, with shared memory (inference).
  • What they want to be known for: AI that works for one person and also becomes part of how the whole company works [yc].
  • Where they sit on the map: L1/L2/L3 = 2/3/1, buyer = sme, business.

4. Three layers vs ours

Layer Them (score + evidence) Us (Marton's logs show)
L1 Company brain 2. They build "shared company memory" from decisions, conversations, documents and activity in connected tools, and agents read it later [yc]. Product Hunt text talks about "shared context" and does not describe memory in detail, so how deep it goes is UNVERIFIED [blog]. A vault of about 1,400 pages, 112 call transcripts and about 8,700 customer quotes, searchable by meaning. Learns from calls and checked work. Built for one business.
L2 Agent orchestration 3. The product is a team of agents. Agents hand tasks to each other on a queue, the owner picks which agents may work together, and risky actions wait for approval [blog] [site]. About 5,600 jobs run by agents from three makers, each checked by a different AI, then automatic tests, before joining the main code.
L3 Execution and learning 1. Agents do the work (tickets, pull requests, reports) and keep an audit trail, but I found no public check of results and no claim that measured results change the next run [blog] [site]. Founder says the aim was agents that "compound instead of reset" [blog], which reads as memory, not measurement (inference). Every failure gets a record and an automatic check (about 160). Offer feedback lines fell from 44 to 18 as rules were added. Only about 15% of jobs reached the main code, and merges were unreliable in October.

5. Same and different

Topic Same or different Them Us
Core belief Same Agents need shared memory and handoffs, not isolated chats [blog]. Same (brief).
Agent handoffs Same Agent-to-agent tasks on a queue, visible on the task [blog]. Marton's logs show a job queue that hands work between agents.
Control Same Permissions, approvals, audit history [blog]. Marton's logs show review by a different AI plus tests, and no-bypass guards.
Buyer Mostly same Small and mid-size teams in many industries, no named customer [site]. Target: SMEs and scale-ups (brief). No customer yet.
Kind of work Different Wide: ops, marketing, support, plus a PM agent and an engineer agent [site] [blog]. One own business so far: coding, content, offers. Candidate: coding and PM work.
How agents improve Different Teach by recording, add or edit skills by hand or by asking the coworker [blog]. Marton's logs show rules added from failures and feedback, each tied to an automatic check.
Proof Different No numbers, no named customers [site]. No customer number yet. Own-business numbers only.
Delivery Different Self-serve software, free tier [site]. The system plus the two founders.
Price Different Per coworker per month, 100 to 550, plus credits [site]. Not set yet.
Stage Different Live product, two Product Hunt launches in 2026, public sign-up [blog] [site]. Nobody but Marton has run it.

6. Wedge overlap

Wedge Fit Why
W1 Trusted 24/7 coding for small startup teams direct Their Sam agent does small site changes, bug fixes and pull requests from Slack, GitHub and Linear, and merges can wait for approval [blog]. Same buyer (small team), same job. They show no 24/7 claim, no review by a second AI, and only small tasks (UNVERIFIED that none exists).
W6 Tool for product managers direct Their Alex agent turns customer feedback into specs, tickets and sprint plans, and chases stalled work [blog]. Same job as "Cursor for product managers", aimed at the team, not the PM (inference).
W2 Seeing whether your agents produce adjacent They give an audit history of what their own agents did [blog]. No view of cost, stuck work or silent failure across other tools (UNVERIFIED that none exists).
W3 Company memory from every call adjacent Both build company memory. Theirs reads connected tools such as Slack, Gmail and Drive [yc]; no call-recording claim found (UNVERIFIED).
W8 The other 90% of an AI skill adjacent Skills library, upload your own, teach by recording, edit from the dashboard or ask the coworker [blog]. Upkeep is left to the user; no claim of automatic improvement from real use (UNVERIFIED that none exists).
W9 Ask my agent before you book my time adjacent Shared memory plus agent handoffs, so a teammate can see why something was done [yc]. Coworkers are roles, not a stand-in for each leader (inference).

7. Whitespace (where they stop)

  • Layer gap: they do the work and log it, but no public check that the result was right and no loop from checked results back into the skill [blog] [site]. To own it we need accepted or rejected work changing a skill, with rework falling over weeks.
  • Number gap: no customer number, no named customer, only mockup dashboards [site] [blog]. To own it we need one workflow run for 4+ weeks with a before/after number the buyer signs.
  • Buyer gap: they say "any team, workflow, or industry" [yc]. A partner will read that as no focus. To own a narrow buyer we need 3 to 5 conversations with one buyer type who say they have the pain.
  • Price and speed: not gaps. Self-serve, a free tier, and 100 to 550 per coworker per month are cheap for a small firm [site]. We should not plan to beat them on price or on time to first use.

8. For the YC application

  • Proven today: Marton's logs show his system runs agents from three makers that do and check work: about 5,600 jobs, each checked by a different AI before automatic tests, and about 160 failure records that each added an automatic check. Backdrop publicly offers human approval and an audit log [blog]. Neither side has a customer number.
  • Draft answer to "how are you different from Backdrop?": DRAFT. "Backdrop lets you build agents and approve what they do. Our system already checks agent work with a second AI and tests, and turns every failure into a rule, every day, on our own business." Founders say it their own way.
  • Future claim (needs proof): "Our agents get better from every piece of work you accept or reject, and we show the number." Needs one customer workflow run for 4+ weeks, a rule-change log tied to checked runs, and rework falling.
  • What we understand that they don't: setup tools (recording, skills library) are not what makes agents trustworthy; checked results are. needs proof on a customer. Marton's logs show it on his own work (offer feedback lines fell from 44 to 18 as 22 rules were added).
  • Copyability (M3): easy. They already hold the inputs: a skills library, edit-by-asking, approvals and a full action log [blog]. Adding "learn from accepted and rejected runs" is a feature for them, it fits their buyer, and it is the kind of change a two-person team ships in weeks (inference). Not hard, so M3 can not score above 3 against Backdrop.
  • Why now (M4): None found in this teardown. Their own relaunch on 15 Oct 2026 is a company change, not a market change [blog]. Do not use "agents can now do the work": that is general AI progress.
  • Do not claim: "shared memory" or "agents that hand work to each other" as our edge (they have both). "AI coworkers" as our headline (their name for the category). "Nobody checks agent work" (we have not shown they do not). "Compounding agents" without a number (their founder uses the idea).

9. Ideas worth taking

  1. Ship a named starting pair (a PM agent and an engineer agent that share context) so a buyer can hire "one" in minutes [blog]. A named entry point is easier to buy than a platform.
  2. Show the handoff on the task itself, so the buyer sees what passed from one agent to the next [blog].
  3. A free tier plus a per-agent monthly price ladder [site] gives a small firm a cheap first step. Test a price in conversation before copying.

10. Unknowns

  • Real traction: users, paying accounts, revenue. Look at their next launch note, or ask a user who has tried the free tier.
  • Funding. One database snippet says $500k pre-seed (Apr 2023), not opened. Check Crunchbase or ask a YC alum.
  • Whether "shared company memory" is a real store that builds over time or context passed inside a conversation. Product Hunt text does not describe memory. Sign up for the free tier and test.
  • Whether the team is still two people. YC shows 2; the site has no team page.
  • Whether anyone on their side measures results or outcomes. No public sign of it.
  • Product Hunt counts are unclear: the page shows 96 upvotes for the 15 Oct 2026 launch, and a search summary gave 158 upvotes and 48 comments for the July launch, while the page itself shows other numbers; I did not resolve which is which.
  • Other funded AI-coworker companies exist in the same space (a search summary named Coworker.ai and Viktor; not opened, UNVERIFIED). A separate teardown is needed before saying how Backdrop ranks among them.
  • A NoCode.Tech article describes Backdrop and compares it to Claude Cowork and ChatGPT Work. I could read only its headline, not the body.

Sources

Technical detail
Slug
backdrop
Lens hash
179096c479
Pass
deep
Stale
no