The week opened with a press release that reframed two months of commentary in a single news cycle.
Anthropic accused DeepSeek, Moonshot, and MiniMax of industrial-scale Claude distillation — twenty-four thousand fake accounts, sixteen million exchanges, with specific allocations across the three labs. The substance of the accusation matters less than the framing: the Western frontier lab is no longer arguing the moat is the model. The accusation makes sense only if the moat is the agent stack — the computer-use traces, the agentic capability layer, the workflow patterns — that the open-weights labs would otherwise have to build from scratch.
The rest of the week confirmed the framing four more ways. Claude Code Remote Control shipped Tuesday and Anthropic disclosed Claude Code at $2.5B ARR. OpenAI launched the Frontier Alliance with four major consultancies the same day. Anthropic acquired Vercept for computer-use Wednesday. Figma shipped OpenAI Codex MCP integration Thursday, eight days after its Claude Code MCP integration. Five separate moves. Same direction. The labs spent the last week of February publicly building the moat above the model.
I am writing this on a Friday twenty-four weeks into building DOS, with the Tuesday evergreen on Skills/Packs/Hooks live alongside this dispatch. The two posts read as the same argument from two angles — the Tuesday essay sketches what an indie's three-layer agent stack should look like; the dispatch records the week the substrate vendors confirmed the layered shape is the actual game.
The week's signal in one sentence
The model has decisively become substrate. The moat is the agent stack — computer-use, MCP integrations, enterprise channel, mobile UX. Indie founders building anywhere in the agent-stack layer just got the strongest validation of the year. Indie founders still building wrappers around raw model APIs just had their moat reframed publicly as someone else's commodity.
The hook: distillation reframes the moat
The single most consequential thing of the five-day window landed Monday morning.
On Mon Feb 23, Anthropic publicly accused DeepSeek, Moonshot, and MiniMax of large-scale Claude distillation (TechCrunch, VentureBeat). Twenty-four thousand fake accounts. Sixteen million exchanges. Specific allocations: MiniMax at 13M (agentic coding and tool-use), Moonshot at 3.4M (agentic reasoning, computer-use), DeepSeek at 150K (alignment and policy probes).
The substance of the accusation is interesting; the framing is structurally important. Anthropic is not saying "they copied our model." Anthropic is saying they harvested our agent traces. The thing being mined is not the weights — it is the operational pattern of how Claude is used as an agent. The complaint only makes sense if Anthropic believes that pattern is the moat.
That confirms the framing every dispatch since November has been arguing toward: the model is increasingly substrate; the value layer is what the model does in agent context. The same week's three other Anthropic moves — Remote Control, Vercept acquisition, $2.5B ARR disclosure — reinforce the framing. Anthropic is investing where the moat is. The accusation against the open-weights labs is partially a defensive announcement and partially a public statement of where the company believes value accrues.
For an indie founder, the implication is direct. Whatever you build above the model is the work the labs are protecting. Whatever you build inside the model is the work the open-weights tier will replicate within a quarter. The architecture argument has a public spokesperson now, and that spokesperson is Anthropic's general counsel.
The cadence intensifies
Three more Anthropic moves rounded out the week, each reinforcing the same direction.
Tue Feb 24 — Claude Code Remote Control + $2.5B ARR (source). Native mobile handoff for terminal coding sessions: QR-code pairing, no port-forwarding or VPN, Max-tier research preview. The $2.5B run-rate disclosure is the second story underneath: Claude Code revenue more-than-doubled since January. The product surface — agent that runs while you sleep, hand off to phone, pick up from desktop — normalizes a UX expectation that did not exist in October.
Wed Feb 25 — Anthropic acquires Vercept (TechCrunch, Anthropic). Seattle-based, ex-AI2 founders, computer-use specialty. Second Anthropic acquisition in three months. Vercept's Vy product sunsets March 25 with a 30-day customer migration. Total raised before acquisition: $50M, with Schmidt, Jeff Dean, Vogt, and Ferdowsi as investors. Anthropic is buying agent-stack capability faster than it is shipping it organically.
The Anthropic week, read end-to-end, says one thing. The lab whose substrate is now the largest enterprise default (Microsoft Copilot from January, ServiceNow, Skills directory) is investing in agent-stack capability above the model across every channel — distillation defense, mobile UX, computer-use acquisition, enterprise distribution. The model is treated as solved enough that Anthropic's strategic spending is going to the layer above.
The OpenAI counter-positioning
OpenAI did not stay quiet.
On Mon Feb 23, OpenAI launched the Frontier Alliance (Fortune, OpenAI). Multi-year enterprise channel with McKinsey, BCG, Accenture, and Capgemini. BCG and McKinsey on strategy; Accenture and Capgemini on systems integration. "Forward deployed engineers" embedded in client work. Direct counter to Anthropic's enterprise-plugin-stack push.
This is structurally identical to Anthropic's enterprise channel pattern — Microsoft Copilot, ServiceNow Build Agent, the Skills directory partner list. The shape is the same: substrate vendor partners with distribution channel that already has the enterprise customer; both sides win by routing through the channel rather than building direct sales.
For an indie founder, the implication ladders up to the question of which channel pattern your product fits inside. If your tool runs as an OpenAI capability inside McKinsey-deployed client work, you are inside the Frontier Alliance distribution. If your tool runs as a Skill inside Claude Code or Microsoft Copilot, you are inside the Anthropic distribution. If your tool runs only on its own surface, you are competing with both partner channels for attention you cannot afford.
The Figma double-tap
The most underrated story of the week is the Figma move.
On Thu Feb 26, Figma shipped OpenAI Codex roundtrip integration via MCP (TechCrunch, OpenAI). Bidirectional flow: Figma Make/FigJam → Codex → code → editable Figma designs.
The integration itself is unsurprising. The date is the story. Figma shipped Claude Code MCP integration eight days earlier, on Feb 18. Eight days later, the parallel OpenAI Codex MCP integration. Figma is shipping parallel integrations to both major coding-agent surfaces in eight-day intervals — through the same MCP protocol, with the same bidirectional contract.
That is the pattern that should reshape every indie tool's distribution thinking. The protocol — MCP — is the boundary. The IDE / surface / chat client is interchangeable on each side of the boundary. If your tool ships an MCP server, it can integrate with whichever coding-agent surface the operator is using, with no per-surface adapter work. The protocol is the new design-system boundary.
For DOS specifically, the Figma double-tap validates the architectural commitment I made in January. MCP is now the layer that determines distribution. Tools that ship an MCP server are vendor-portable across the coding-agent surfaces. Tools that do not are visible only to one surface at a time.
The pattern: stack assembled in public, on enterprise rails
The five-day window in one frame
- Frontier model lead as the durable moat
- Open-weights catch-up as the existential threat
- Indie tooling as a wrapper around model APIs
- Distribution as a "we'll figure it out later" problem
- The lab competes on benchmarks
- Distillation accusation reframes the moat as agent traces, not weights
- Anthropic invests in mobile UX, computer-use acquisition, enterprise distribution
- OpenAI bolts four consultancies onto its Frontier platform
- Figma ships parallel MCP integrations eight days apart, treating MCP as the boundary
- Indie tooling ships MCP servers and rides inside enterprise channel default surfaces
If you wanted evidence for the year's prevailing narrative — "the substrate is plural, the protocols are public goods, the workflow layer is the moat, the rules layer is being codified, the hyperscalers are vertically integrating, the flagship is agentic-coding-shaped, the substrate has bifurcated, agentic primitives commoditize into the free tier" — this week added the ninth refinement: the moat is being assembled in public, on enterprise rails, above the model layer. The labs themselves are now publicly arguing for the framing this dispatch series has been writing toward since November.
Two angles for an indie founder
What an indie founder building on the substrate should do this week
- Treat MCP as the design-system boundary, not as an integration. The Figma double-tap (Claude Code MCP on Feb 18, Codex MCP on Feb 26) proves the protocol is the actual contract. If your tool ships an MCP server, you are vendor-portable across coding-agent surfaces — the cost is one server, the benefit is distribution across however many MCP-aware surfaces the operator runs. If your tool requires per-surface adapters, you are investing in a narrowing market. Build to MCP first. The integration with any specific surface is the second-order work.
- Match the "agent that runs while you sleep" UX expectation. Claude Code Remote Control normalizes mobile-handoff of long-running agent sessions. By Q2, every credible coding-agent surface will ship some version of this. Indie tooling that still assumes "developer at laptop" is about to feel dated. Worth thinking about how your product surfaces in-flight work to a phone, a notification, an asynchronous resume — even before the demand is loud. The ones who position early ride the demand wave; the ones who wait fight against the new expectation.
- Pick your enterprise channel deliberately, or accept that you have none. OpenAI's Frontier Alliance with McKinsey/BCG/Accenture/Capgemini is the third major channel pattern of Q1 (after Microsoft Copilot and ServiceNow Build Agent). Indie tooling that does not ship inside one of these channels is competing with their internal stacks for attention. Pick the channel whose customer matches yours, ship as a Skill or MCP server inside that channel's preferred surface, and the customer-acquisition story compresses by months. Pick none, and the customer-acquisition story is yours alone — which has worked for some founders, but rarely as cheaply as the channel-route does.
What this changes for DOS
Two design decisions hardened this week, both of them coming out of the moat-above-the-model framing rather than any single news item.
One. The MCP server for DOS's Sentinel pack (sketched in W8, still pending production ship) becomes a Q1 commitment, not a Q2 deferral. The Figma double-tap proved that MCP is the boundary that determines distribution. Sentinel as a Claude-Code-only Skill ships to Claude Code's surface; Sentinel as an MCP server ships to whatever the operator runs (Claude Code, Cursor, Codex, Aider). The cost of the MCP server is roughly two days. The cost of skipping it is everything else.
Two. Studio's roadmap for mobile-aware long-running session UX moves from "interesting, future" to "scoped this quarter." The Claude Code Remote Control launch is the demand signal. The first DOS user who routes a long-running agent task through Studio and wants to know "did it finish?" from their phone is not far away. The shape of the eventual surface is unclear; the architectural prep — exposing in-flight task state via Studio's gateway — is not. Building the prep now means the actual surface ships when the demand surfaces.
That is the kind of forcing function the moat-above-the-model week is structurally designed to produce. The framing the substrate vendors made public this week is the framing this dispatch series has been making private since November. Now both sides agree, and the work follows the agreement.
What I am watching for next week
The thread that runs through the week
The model has decisively become substrate. The moat is the agent stack — distillation defense, mobile UX, enterprise channel, MCP integrations. Anthropic invested in all four directions in five days. OpenAI invested in the consultancy channel. Figma invested in MCP as the boundary. The labs are publicly building the moat above the model that this dispatch series has been arguing for since November.
For an indie founder, the playbook reduces. Ship MCP first. Position inside an enterprise channel, or accept that you have none. Match the mobile-handoff UX expectation before it becomes table stakes. Bet the moat on the operator loop on top — context, memory, conventions, failure pipelines, the convention catalogs the agent reads at session start — because that is the only piece of the stack neither the labs nor the open-weights tier can vertically commoditize.
That is what Studio is becoming. The moat-above-the-model week made the strategic argument operational, and the Sentinel MCP server moves up the queue accordingly.
— Lucas
Sources verified the week of Feb 23-27, 2026: Anthropic distillation accusation — TechCrunch (Feb 23) · VentureBeat on distillation (Feb 23) · Claude Code Remote Control — VentureBeat (Feb 24) · OpenAI Frontier Alliance — Fortune (Feb 23) · Anthropic acquires Vercept — TechCrunch (Feb 25) · Figma + OpenAI Codex MCP — TechCrunch (Feb 26)
Was this page helpful?