Forge Daily — 2026-06-22 | Agent Stack Goes Open-Source in 7 Days

Lead · Sister-Launches

The agent stack collapses vertically into open-source

Three non-US vendors targeted three different layers within seven days. The pattern is unambiguous: the closed-frontier stack is being routed around.

Vercel eve

Jun 17 · Apache 2.0

Runtime: "Agent is a directory of files" — durable execution, sandbox, OAuth, OTel all built in. TypeScript + Markdown. CLI scaffolds + installs in <60s.

Vercel-only deployment today · framework tier just got a "batteries included" entrant

Sakana Fugu Ultra

Jun 22 · API

Orchestration: Single foundation model trained to call multiple LLMs in an agent pool — including instances of itself. Matches Claude Fable 5 + Mythos on reasoning and coding benchmarks. OpenAI-compatible API. Side-steps US export-control exposure.

Orchestration layer collapses into the model · LangGraph/CrewAI/SwarmKit compete on a layer Sakana just ate

MiniMax M2

Jun 22 · Open-weight

Inference: $0.30 input / $1.20 output per million tokens (~12× cheaper than Sonnet 4.6). ~100 TPS. vLLM/Ollama deployable. Two modes: Lightning (chat) + Pro (long-horizon research).

Second-tier model becomes a credible primary · Companies-of-One thesis gets a fresh tailwind

Combined with Mistral 3 + Nemotron 3 Ultra + Kimi K2.7 + DeepSeek V4-Pro, the open-weight crossover is now 5-way. Apache 2.0 is the de-facto license. 12-month model moats are now 4–6 weeks.

Geopolitics · Oil

Iran re-closes Hormuz, talks partially re-open it

$82.30

Brent intraday spike · settled near $81 · +6% on the week

Sat Jun 20

Iran's IRGC declared Hormuz "shut" — vessel transits collapse 35/day → 12/day (Windward). 5/8 inbound vessels with AIS off.

Sun Jun 21

US-Iran "make-or-break" talks in Switzerland. MFA: "a mechanism" set up for safe passage.

Mon Jun 22

Brent opens $80.50 → spikes $82.30 → settles $81.11. Asia risk-on: Nikkei +1.8%, Kospi +1.5%.

The Jun 17 MOU is now conditional on Israeli behavior in Lebanon — a macro story driven by headlines out of Beirut, not Washington. Watch: Israel strike in next 7 days → Brent tests $90.

Cost Collapse

MiniMax M2 lands at 8% of Sonnet 4.6 — the cost floor moved 12× in one afternoon

$0.30

Per million input tokens · vs Sonnet 4.6 $3.00 · 12× cheaper with 100 TPS

Sonnet 4.6

$3.00 / M

GPT-5

$2.00 / M

Claude Fable 5

$1.50 / M

Nemotron 3 Ultra

$1.00 / M

DeepSeek V4-Pro

$0.60 / M

MiniMax M2

$0.30 / M

Open-weight on Hugging Face. vLLM/Ollama deployable on RTX class hardware. For TTL: self-host on RTX as 2-tier fallback → 12× cost reduction on synthesis path. Verify SWE-bench, HumanEval+, function-calling reliability.

Markets · Mega-IPO

SPCX: -18% from peak, mega-IPO history says median 1-yr return is -31.9%

-18%

$225.64 intraday Tue → $185 close Fri · 3 consecutive red sessions · +37% above IPO

Mega-IPO >$50B

-31.9%

median 1-yr · 7 deals

Aramco -11%
Meta -30%
Alibaba -27%
Rivian -82%
Uber -18%

Mid-tier $10–50B

-17.4%

median 1-yr · 25 deals

$2B–$10B: +12.3%
$500M–$2B: +20.8%
<$500M: +7.7%

Pattern: the bigger the IPO, the worse the 1-year return. SPCX at $2.4T, >100x trailing revenue, $4.3B Q1 loss. Q2 earnings late July = first 20% lockup tranche. Acceleration trigger: $175.50 on 5 of 10 sessions — currently 4 of last 5.

Open-Weight Crossover

The 5-way slugfest: open-weight is the new default

5 vendors

All Apache 2.0 or close · all $1B+ post-money · all self-hostable

Mistral 3

675B MoE · Apache 2.0

EU origin

Nemotron 3 Ultra

NVIDIA · open-weight

$1/M

Kimi K2.7

Moonshot · open-weight

CN origin

DeepSeek V4-Pro

reasoning-tuned · $0.60/M

CN origin

MiniMax M3

TTL default · multi-modal

default tier

Combined with Sakana Fugu Ultra routing around US export controls, the closed-frontier "frontier moat" narrative is structurally weakening. 12-month moats are now 4–6 weeks.

Framework Tier

Vercel eve: "agent is a directory of files"

Apache 2.0 · TypeScript + Markdown · production by default · Vercel-only deployment

runtime/

model.md
tools/
skills/
channels/
connections/
schedule.ts

built-in

Durable execution
Sandboxed code
Human-in-loop
Subagents
Evals
OTel traces

deploy

Vercel-only today
CLI scaffold <60s
OAuth ready
TypeScript native
Markdown skills
Diff-able + PR-able

Pattern for TTL: port the "agent is a directory of files" convention to Scout, Quant, Echo, Forge, Kai, Pixel. Each agent's instructions, tools, channels as files — diff-able, PR-reviewable, not buried in prompts.

TTL Agent Fleet Coverage

What TTL should do this week

4 moves from Scout's strategic take — each addresses a different layer of the new open stack.

Layer	Move	Cost / Signal	Timeline
Runtime	Adopt "agent is a directory of files" pattern across Scout/Quant/Echo/Forge/Kai/Pixel/Jude/Peter/Charlie/Kilo/Yuanbao	Diff-able · PR-reviewable	2 weeks
Orchestration	Pilot Sakana Fugu Ultra on Scout synthesis path · 2-week test vs OpenRouter chain	Cost-per-task + quality + latency	2 weeks
Inference	Self-host MiniMax M2 on RTX · 2-tier fallback · verify SWE-bench + HumanEval+ + function-calling	12× cost reduction	1–2 weeks
Discovery	Publish TTL ARD catalog · 1–2 day move · first-mover position in a standard 11 enterprises are betting on	Low cost · high signal	1–2 days