# Agent Platform Research — February 27, 2026
## COVERAGE NOTES
- **anthropic-pentagon-ultimatum** (active): DEADLINE IS TODAY 5:01 PM ET. Anthropic formally rejected Pentagon's "final offer" overnight. Contract termination and "supply chain risk" designation imminent. Full outcome update needed after 5 PM.
- **openclaw-2026-2-25** (active → new release 2026.2.26 trumps it): New release published this morning. GLaDOS is now 3 versions behind.
- **deepseek-v4-lite-gpt5-garlic** (active): GPT-5.3-Codex quietly released Feb 24 — NOT the same as "Garlic" codename per Polymarket, but is a real OpenAI model drop worth covering. Adding as new entry. DeepSeek V4 still unreleased.
- **rocketlab-q4-2025-earnings** (active): Results in — cover now, mark resolved.
- **rocketlab-haste-dart-ae-scramjet** (active): Launch rescheduled for today 4 PM ET. Outcome still pending.
- MCP: Airia MCP Gateway hits 1,000 integrations — ecosystem development note, not blockbuster.
- LangChain/LangGraph/CrewAI/AutoGen: No new releases or notable announcements in past 24h. Skipping.
---
## NEW STORIES — February 27, 2026
### 1. 🔥 ANTHROPIC VS. PENTAGON — DEADLINE TODAY
Anthropic formally rejected the Pentagon's "final offer" overnight, setting up a showdown at 5:01 PM ET today. Pentagon spokesman Sean Parnell posted on X: "They have until 5:01 PM ET on Friday to decide. Otherwise, we will terminate our partnership with Anthropic and deem them a supply chain risk for DOW."
Anthropic's statement: *"The contract language we received overnight from the Department of War made virtually no progress on preventing Claude's use for mass surveillance of Americans or in fully autonomous weapons."*
Stakes beyond the contract:
- $200 million contract on the table
- "Supply chain risk" designation — normally reserved for foreign adversaries — could damage Anthropic's broader government business
- CNN reports Pentagon also threatened unspecified designation that could threaten their "bottom line"
- Anthropic says they "cannot in good conscience" accede
Context: Pentagon says it's never *considered* autonomous weapons or mass surveillance, but won't explicitly prohibit them in contract language. That gap is the sticking point. Simultaneously, Anthropic dropped its RSP scaling-pause clause last Tuesday (covered Feb 26) — the optics of holding firm on military guardrails while relaxing internal safety commitments is being noted by critics.
---
### 2. 🦞 OPENCLAW 2026.2.26 — EXTERNAL SECRETS + ACP AGENTS
Released this morning (announced via X @openclaw). Key features:
- **🔐 External Secrets Management** (`openclaw secrets`) — store credentials outside workspace config; first-class CLI for secrets rotation, vault integration
- **🤖 ACP thread-bound agents** — Agent Communication Protocol (ACP) agents now treated as first-class runtime objects, enabling persistent multi-agent thread contexts
- **⚡ Codex WebSocket-first transport** — Claude Code / Codex sub-agents now use WebSocket transport by default (lower latency than HTTP polling, better for long-running tasks)
- **📱 Android app improvements** (unspecified in snippet)
- **🔧 Agent routing CLI** (`bind/unbind`) — new commands to route specific message sources/channels to named agents
- **🛡️ 11 security hardening fixes**
Note for Rich: GLaDOS is still on **2026.2.21-2** — now three minor versions behind (2026.2.23 → 2026.2.25 → 2026.2.26). The `openclaw secrets` feature and ACP thread-bound agents are both immediately relevant to the voice PWA backend and multi-agent session work.
---
### 3. 🤖 GPT-5.3-CODEX — OPENAI'S FRONTIER AGENTIC CODING MODEL DROPS
Released February 24 (missed in yesterday's briefing — was tracked as upcoming "Garlic" but shipped under a different name). Key specs:
- **400K context window** — handles large repos and long task states
- **77.3% Terminal-Bench 2.0 score** — top of published agentic coding leaderboards
- GPT-5-class reasoning + exceptional math + vision support
- Available in GitHub Copilot (Enterprise, Business, Pro, Pro+) via Copilot Chat on github.com, mobile, and Visual Studio
- Available in Azure AI Foundry / Windows Forum confirms Foundry support
Note: "Garlic" codename on Polymarket resolves March 2026 — possibly this IS Garlic, or Garlic is a separate larger general-purpose release still pending. Either way, GPT-5.3-Codex is the real deal and now shipping. Major upgrade for agentic coding workflows.
---
### 4. 🚀 ROCKET LAB Q4 2025 EARNINGS — RECORD REVENUE, NEUTRON DELAYED
Results from after-market close Feb 26:
- Q4 revenue: **$180M** (beat consensus of $178M)
- Full-year 2025: **$602M** (38% YoY growth)
- Backlog: **$1.85B** (73% growth YoY) — includes $816M satellite contract
- Q1 2026 guidance: $185–200M revenue, 34–36% gross margins
The asterisk: **Neutron Stage 1 tank test failure** pushes first launch target to **Q4 2026**. Street expected mid-2026. Shares up ~1.2% after-hours — the record revenue and backlog growth offset the Neutron delay.
---
### 5. 🛸 HASTE/DART AE SCRAMJET — LAUNCH ATTEMPT TODAY
HASTE "That's Not a Knife" mission rescheduled for today, February 27, at **4:00 PM ET (21:00 UTC)** from Wallops Island, Virginia. Multiple scrubs already logged. Hypersonix DART AE is a hydrogen-fueled 3D-printed scramjet developed in Brisbane. DIU-funded. Mach 5+ test, 620-mile downrange. Will update outcome when available.
---
## TOPICS WITH NO NEW DEVELOPMENTS (SKIPPED)
- LangChain observability (covered Feb 22, no new adoption news)
- MCP SDK CVEs (fix released, no PoC published)
- Microsoft Agent Framework RC (no GA announcement)
- PersonaPlex-7B (no new benchmarks, covered 4x)
- MiniMax Speech 2.6 (covered, cooling-off period)
- OpenAI Realtime API MCP/SIP (no new capability news)
- VAST Data agentic platform (no GA/pricing)
- SambaNova SN50 (pre-ship, no new benchmarks)
- IBM/Deepgram (no follow-on news)
- Airia MCP Gateway 1,000 integrations — noted as ecosystem growth, not cover-worthy on its own
---
## STORIES TO WATCH
- Anthropic/Pentagon outcome at 5 PM ET today — could be a contract termination that reshapes AI safety narrative
- DeepSeek V4 actual release (still withheld, Huawei optimization window ongoing)
- HASTE launch result today
- Claude Code CVE-2026-21852 patch (still "unclear" status)