AI agent & dev infrastructure

Thesis

The agent economy crossed from demo into production infrastructure in 2024–26. Three signals: (1) Anthropic's Model Context Protocol (MCP) shipped 25 Nov 2024, was adopted by OpenAI on 26 Mar 2025 (Agents SDK) and Google DeepMind in April 2025, and was donated to the Linux Foundation's Agentic AI Foundation (AAIF) on 09 Dec 2025 — governance is now multi-vendor, not Anthropic-owned. (2) Anthropic reached a ~$30B annualized run-rate by April 2026, up from the $14B disclosed at the Series G close on 12 Feb 2026 ($30B raised at $380B post-money); OpenAI closed the largest private round in history on 31 Mar 2026 — $122B at $852B post-money, with $2B/month revenue (~$24B annualized) and 900M+ ChatGPT weekly actives. (3) Q1 2026 was the largest VC quarter on record at $285.5B (CB Insights); the four largest rounds (OpenAI $122B, Anthropic $30B, xAI $20B, Waymo $16B) absorbed 65% of the quarter.

When agents reach production, a new primitive layer emerges: context management, evaluation frameworks, guardrail runtimes, agent orchestration, observability. This is the devtools layer — analogous to what Datadog, LaunchDarkly, and Segment built for the cloud era. The market is global by default; SEA founders have a cost-to-build and time-zone advantage against North American and European builders.

Signal we want to back

• Agent-native, not LLM-wrapper: the product assumes it's serving production agent workloads (long-running, non-deterministic), not a chat UI for humans.
• Developer-first distribution: GitHub stars, npm downloads, Discord engagement — proof that developers are actively adopting. Bottom-up, not enterprise sales.
• Deeply technical founder with taste: has shipped real production agents (not just prototypes). GitHub history matters more than résumé.
• Fits a protocol/standard (MCP, OpenAPI, Google's A2A): accepts that agents operate inside the foundation-model ecosystem, not against it.
• Revenue model matches usage: usage-based pricing (per invocation, per token, per success) — not per-seat for developers.

Signal we skip

• Vertical GPT wrappers: "ChatGPT for lawyers," "ChatGPT for doctors." Margin gets squeezed the moment OpenAI/Anthropic ships a general agent for those domains.
• No-code agent builders for consumers: small market, high CAC, low retention — already proved out via Zapier AI / Pipedream AI wave 2023–24.
• Pure AI SDKs: Vercel AI SDK, Mastra, LangChain core are all free. A company needs a layer on top of the SDK (hosting, eval, state) to have a moat.
• New foundation models without a data moat: no practical reason to back a generic new frontier model at this point, unless there's unique proprietary data (medical, legal archive, industrial telemetry).

Recent data

• Anthropic economics: Series G closed 12 Feb 2026 at $380B post-money, $30B raised; disclosed run-rate $14B (up from ~$1B in Jan 2025, "over 10× annually for three years"). By April 2026, Bloomberg/Sherwood put run-rate at ~$30B [PYMNTS]. Claude Code ARR $2.5B, doubled since 01 Jan 2026; enterprise >50% of Claude Code revenue. 8 of the Fortune 10 are customers; 1,000+ customers spending >$1M/year (doubled in under 2 months). ~4% of all GitHub public commits are now authored by Claude Code (doubled MoM).
• OpenAI economics: $122B at $852B post-money, closed 31 Mar 2026 — Amazon $50B (of which $35B contingent on IPO-by-2028 or AGI milestone), Nvidia $30B, SoftBank $30B. $2B/month revenue (~$24B annualized); 900M+ ChatGPT WAU, 50M+ paid subs; 92% of Fortune 500 use ChatGPT; 7M+ workplace seats. Stargate JV targets up to $500B over four years.
• Frontier lineup (24 Apr 2026): Claude Opus 4.7 (16 Apr 2026, 1M context, $5/$25 per 1M, new tokenizer); Sonnet 4.6 (17 Feb 2026, $3/$15, 1M ctx); Haiku 4.5 ($1/$5, 200K). GPT-5.4 (05 Mar 2026, $2.50/$15, 1M ctx); GPT-5.5 shipped 23 Apr 2026 to ChatGPT/Codex only, API deferred, announced at $5/$30. Gemini 3.1 Pro (19 Feb 2026, $2/$12 up to 200K, $4/$18 beyond); Gemini Pro paid-only from 01 Apr 2026. Mistral Large 3 (Dec 2025, open-weights MoE); DeepSeek V3.2 ($0.28/$0.42); Grok 4.20 (31 Mar 2026, 2M ctx, $2/$6).
• Protocols: Anthropic donated MCP to AAIF/Linux Foundation on 09 Dec 2025 — founding Platinum members include Anthropic, OpenAI, Google, Microsoft, AWS, Bloomberg, Cloudflare, Block. Google's Agent-to-Agent (A2A) launched 09 Apr 2025, donated to Linux Foundation 23 Jun 2025; >150 supporting organizations by Apr 2026. Public MCP server counts span ~12K–22K depending on registry (mcp.so 20,354; Glama 21,586); SDK downloads ~97M/month (vendor estimate, Taskade Apr 2026).
• Agent devtools M&A + mega-rounds: Anysphere/Cursor Series D $2.3B at $29.3B (13 Nov 2025; April 2026 talks at ~$50B ⚠️ not closed). Cognition (Devin) Series D ~$400M at $10.2B (Sep 2025; April 2026 talks at ~$25B ⚠️ not closed). LangChain Series B $125M at $1.25B (20 Oct 2025). Braintrust Series B $80M at $800M (17 Feb 2026). M&A: CoreWeave → Weights & Biases ~$1.7B; Palo Alto Networks → Protect AI >$500M; ClickHouse → Langfuse (Jan 2026).
• VC context: Global AI VC 2025 = $226B (CB Insights, 48% of total venture) or $270.2B (PitchBook, 52.7%) — use both, never average. Q1 2026 = $285.5B global (largest quarter ever); Crunchbase puts $242B = 80% of Q1 going to AI. Carta Q4 2025 seed post-money median $24M (+33% YoY); Series A $78.7M (+37%); median seed→A wait 616 days (record).

Portfolio stance

Fund I plans 0–2 deals here — globally sourced, no SEA requirement. Fit: companies with SEA-origin founders (VN, SG, diaspora) building for a global market. Ticket: $100–200K. Willing to co-invest with devtools-focused funds (Costanoa, Amplify, Insight).

Pipeline currently has 2 companies: an evaluation framework (SG team, US customers) and a context-memory layer (VN diaspora team, open-source traction).

What changed this quarter

• Frontier model cadence stayed aggressive: Claude Opus 4.7 (16 Apr 2026) + GPT-5.5 (23 Apr 2026, one day before this memo) + Grok 4.20 (31 Mar 2026) + Gemini 3.1 Pro paid-only (01 Apr 2026). Two practical implications: (a) agent-devtools pricing has to assume unit costs keep falling; (b) "Anthropic-only" or "OpenAI-only" product positioning is increasingly fragile — multi-model routing is table stakes.
• Q1 2026 was the largest VC quarter in history: $285.5B global, with the four top AI rounds absorbing 65%. Implication for Seed/Pre-A devtools: mid-market valuations got dragged up (Carta Q4 2025 medians: Seed $24M, Series A $78.7M). Entry discipline matters more, not less — we keep $100–200K tickets and skip "AI premium" rounds.
• Protocol governance de-risked: MCP is now Linux Foundation-stewarded; A2A has >150 supporters. Devtools built on MCP are now a bet on a standard, not a vendor. Hugging Face MCP Server (Jun 2025) + independent registries (mcp.so, Smithery, PulseMCP) form a competing distribution layer — new monetization opportunity.
• SEA angle: we're tracking two teams with VN/SG founders building MCP-native tooling (evaluation + context persistence). If terms fit, one may be Fund I's first check in this space in Q3.