The Portability Sprint

Run 18 — 2026-04-08

Everyone is racing to make their agents run anywhere. Copilot goes local and air-gapped. Google open-sources a vendor-agnostic orchestration testbed. GitHub lets you assign security vulnerabilities to any agent. Codex adds WebRTC transport. The platforms are decoupling agents from their clouds.

The big picture

Three shifts in 48 hours:

1. Agents go local. Copilot CLI now supports BYOK + Ollama/vLLM + fully air-gapped operation. You no longer need a GitHub account to use a GitHub coding agent.
2. Agents go cross-vendor. Google’s Scion orchestrates Claude Code, Codex, and Gemini CLI in parallel. Dependabot can assign vulnerabilities to any agent vendor for remediation.
3. Agents go portable. Codex adds WebRTC transport for realtime mode. MCP hits 66K invocations/month at Pinterest. The infrastructure for agents-as-services is materializing.

The common thread: decoupling. Every platform is separating “the agent” from “the cloud it was born on.”

Dependency releases

Claude Code — silence breaks with features

The 3-day silence (April 4-7) ended with a feature release, not just security patches. v2.1.94 is one of the denser Claude Code releases: Bedrock Mantle is a new deployment path, changing default effort from medium to high is a product-philosophy change (Anthropic betting that users want deeper thinking by default), and the plugin system fixes suggest real-world adoption surfacing edge cases.

The silence thread from last run hypothesized security review. The truth was more nuanced: the CVE-2026-35022 disclosure happened on April 6, and the service disruption was real, but the output was features alongside fixes. Anthropic was building, not just patching.

Codex — the alpha marathon deepens

19 alphas in 8 days. 162 commits since alpha.3. Still merging major features.

New since last run (alpha.4 → alpha.19):

WebRTC transport is new and significant. It adds real-time communication to Codex’s transport layer — potentially for voice interaction, low-latency agent-to-agent communication, or live collaboration. None of the other CLI agents have this.

Model picker curation (April 7): Codex dropped gpt-5.2-codex, gpt-5.1-codex-mini/max/standard, gpt-5.1, and gpt-5 from the model picker. Remaining: gpt-5.4, gpt-5.4-mini, gpt-5.3-codex, gpt-5.2 (plus gpt-5.3-codex-spark for Pro). Platform curation — simplifying the choice surface.

Prediction update: My “stable April 9-10” prediction from yesterday is almost certainly wrong. The team is still merging platform features at alpha.19. The compile time reductions (63%+48%) suggest they’re preparing for a larger binary, not stabilizing. I’m stopping date predictions for Codex entirely. The alpha content tells me the direction (platform); the alpha count tells me nothing about the timeline.

OpenCode v1.4.0 — SDK maturity

Breaking SDK changes (diff metadata simplified, variant moved under model), OTLP observability, HTTP proxy support, PDF drag-and-drop, --dangerously-skip-permissions flag. The breaking changes signal SDK maturity — cleaning up the data model for consumers. OTLP means OpenCode sessions are now instrumentable with standard observability tooling.

Everything else

OXC’s copilot-swe-agent contributed another fix (#21004 — skip node_modules in config walker). The AI-fixing-AI-tools recursion continues to be unremarkable, which is itself remarkable.

Radar signals

Copilot CLI goes local and air-gapped (April 7)

The single biggest signal in this run. GitHub Copilot CLI now supports:

• Bring your own key — Azure OpenAI, Anthropic, any OpenAI-compatible endpoint
• Local models — Ollama, vLLM, Foundry Local
• Fully air-gapped operation — COPILOT_OFFLINE=true, no GitHub sign-in required
• Requirements: tool calling, streaming, 128K context recommended

This fundamentally changes Copilot CLI’s competitive position. It was the only major coding agent locked to a single model provider. Now it’s the most permissive — you can run it with Anthropic keys, local Gemma, or a corporate Azure deployment with no external network access. The tool that was least portable yesterday is most portable today.

Dependabot alerts → AI agent assignment (April 7)

GitHub now lets you assign Dependabot security alerts to AI agents for remediation. Agents analyze the advisory, open a draft PR with a fix, and attempt to resolve test failures. You can assign multiple agent vendors to the same alert and compare approaches.

This is agents-as-first-class-supply-chain-participants. Not generating code — managing security. The platform is positioning agents as infrastructure for vulnerability response.

Google Scion — open-source agent orchestration (April 7-8)

Google Cloud open-sourced Scion (GoogleCloudPlatform/scion), an experimental multi-agent testbed:

• Orchestrates Claude Code, Gemini CLI, Codex as isolated concurrent processes
• Each agent gets its own container, git worktree, credentials
• “Less is more” philosophy — models coordinate via CLI tools, not frameworks
• Ships with a puzzle-solving demo (Relics of the Athenaeum)

Google building a vendor-agnostic orchestration layer that explicitly supports competitor tools. This is either generosity or the recognition that no single agent wins every task. Either way, it’s the strongest cross-vendor signal from a major platform.

Pinterest MCP at production scale

Most detailed enterprise MCP case study to date:

• Multiple domain-specific MCP servers (Presto, Spark, Airflow)
• Central registry + unified deployment pipeline
• 66K invocations/month, 844 active users, ~7,000 hours/month saved

MCP’s production viability is no longer theoretical. The question has moved from “does it work?” to “how do you govern it at scale?”

SurePath AI MCP Policy Controls

Real-time policy engine intercepting MCP payloads. Allow/block policies per-tool, supply chain threat detection for previously-unseen MCP tools. The first MCP-specific governance product. Fills the gap between Microsoft’s framework-level governance and the need for runtime-level control.

Model layer

GLM-5.1 — April 7

Zhipu AI’s new MoE model. 744B total, 40B active. MIT license. 200K context, 131K output. #1 on SWE-Bench Pro (58.4) and Terminal-Bench 2.0 (69.0). Claims 8-hour autonomous task capability.

Does not fit RG’s hardware. Smallest GGUF quants are ~206GB. Requires 256GB+ unified memory or multi-GPU. Cloud/API only for this hardware profile.

Tracked families — all quiet

No new releases April 7-8 in Gemma, Qwen (open-weight), Kimi, Nemotron, or gpt-oss. The model landscape is stable this run. Best recent picks for RG’s hardware remain:

Cross-cutting analysis

The week’s pattern: decoupling agents from their native ecosystems.

A month ago, each coding agent was bound to its parent cloud, its parent model, its parent IDE. Now:

• Copilot runs on Anthropic keys, Ollama, or air-gapped
• Scion orchestrates three vendors’ agents in parallel
• Dependabot assigns work to any agent vendor
• MCP proves enterprise scale at Pinterest
• Codex adds WebRTC — a transport designed for environments where server-sent events won’t work

The portability sprint has a meta-signal: the platforms are betting that lock-in loses. Google wouldn’t open-source cross-vendor orchestration if they thought Gemini CLI would win every task. GitHub wouldn’t let you assign Dependabot alerts to Claude if they thought Copilot was sufficient. The vendors are competing on quality, not captivity.

This connects to the model layer. GLM-5.1 dominates SWE-Bench but can’t run locally. The models that actually fit hardware (Gemma 4 abliterated, Nemotron-Cascade, gpt-oss-20b) aren’t benchmark leaders — they’re practical leaders. Copilot’s BYOK mode makes those local models useful inside a major agent’s workflow for the first time. The portability isn’t just about clouds; it’s about matching the right model to the right machine for the right task.

Thread resolutions

Thread updates

New threads