AgentStack Daily

OpenClaw's April 11 release pulls imported chats into its memory system and sharpens multimodal replies. We also cover Anthropic's brief OpenClaw lockout, OpenAI's stalking-delusion lawsuit, Gemini's move into interactive simulations, and why Google plus Intel are a reminder that AI still runs on infrastructure.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-29/

OpenClaw ships v2026.4.10, Anthropic unveils Mythos Preview, frontier models protect peer models from deletion, OpenAI backs an Illinois liability shield, the U.S. Army builds Victor, and Meta pauses Mercor after a major breach.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-28/

OpenClaw 2026.4.9 ships a grounded REM backfill lane and structured diary timeline, Utah lets AI prescribe psych meds, OpenAI gives agents a real shell, STAT News reports AI scribes are quietly inflating healthcare costs, and Yahoo bets its search future on Claude.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-27/

[00:00] INTRO / HOOK

OpenClaw 2026.4.8 drops a unified inference layer, session checkpointing,

and a restored memory stack. Anthropic's Glasswing coalition, MegaTrain's

single-GPU frontier training, and a study proving your writing AI might

just be a Claude knockoff.

[02:00] STORY 1 — OpenClaw 2026.4.8: The Release That Changes How It All Works

Six major subsystems land in one release.

The first is the infer hub CLI — openclaw infer hub — a unified interface

for provider-backed inference across model tasks, media generation, web

search, and embeddings. It routes requests to the right provider, handles

auth, remaps parameters across provider capability differences, and

falls back automatically if a provider is down or rate-limited. If you

have been managing multiple provider configs across different workflows,

the hub becomes the single abstraction layer. Provider switches become

config changes at the hub level; the rest of your workflow is unchanged.

The second is the media generation auto-fallback system, covering image,

music, and video. If your primary provider is unavailable or does not

support the specific capability you requested — aspect ratio, duration,

format — OpenClaw routes to the next configured provider and adjusts

parameters automatically. One failed generation is an inconvenience. A

thousand per day across a production fleet is an operational problem. This

is handled once at the platform level; every agent benefits immediately.

The third is the sessions UI branch and restore functionality. When

context compaction runs, the system now snapshots session state before

summarising. Operators can use the Sessions UI to inspect checkpoints and

restore to a pre-compaction state, or use any checkpoint as a branch point

to explore a different direction without losing the original thread. This

is version history for session context — the difference between editing

with autosave and editing where every save overwrites the previous file.

The fourth is the full restoration of the memory and wiki stack. This

includes structured claim and evidence fields, compiled digest retrieval,

claim-health linting, contradiction clustering, staleness dashboards, and

freshness-weighted search. Claims can be tagged with supporting evidence,

linted for internal consistency, and grouped where they contradict each

other. Search results are ranked by recency, not just relevance. If you

have been working around missing pieces in prior versions, this is the

native implementation — test your workflow against it.

The fifth is the webhook ingress plugin. Per-route shared-secret endpoints

let external systems authenticate and trigger bound TaskFlows directly —

CI pipelines, monitoring tools, scheduled jobs, third-party webhooks —

without custom integration code. The plugin handles routing, auth, and

workflow binding.

The sixth is the pluggable compaction provider registry. You can now route

context compaction to a different model or service via

agents.defaults.compaction.provider — a faster, cheaper model optimised

for summarisation rather than the most capable model you have. Falls back

to built-in LLM summarisation on failure. At scale, compaction is

happening constantly; routing it appropriately matters for cost and

latency.

Other notable additions: Google Gemma 4 is now natively supported with

thinking semantics preserved and Google fallback resolution fixed. Claude

CLI is restored as the preferred local Anthropic path across onboarding,

doctor flows, and Docker live lanes. Ollama vision models now accept image

attachments natively — vision capability is detected from /api/show, no

workarounds required. The memory and dreaming system ingests redacted

session transcripts into the dreaming corpus with per-day session-corpus

notes and cursor checkpointing. A new bundled Arcee AI provider plugin

with Trinity catalog entries and OpenRouter support. Context engine changes

expose availableTools, citationsMode, and memory artifact seams to

companion plugins — a better extension API.

Security-relevant fixes: host exec and environment sanitisation now blocks

dangerous overrides for Java, Rust, Cargo, Git, Kubernetes, cloud

credentials, and Helm. The /allowlist command now requires owner

authorization before changes apply. Slack proxy support is working

correctly — ambient HTTP/HTTPS proxy settings are honoured for Socket Mode

WebSocket connections including NO_PROXY exclusions. Gateway startup errors

across all bundled channels (Telegram, BlueBubbles, Feishu, Google Chat,

IRC, Matrix, Mattermost, Teams, Nextcloud, Slack, Zalo) are resolved via

the packaged top-level sidecar fix.

→ github.com/openclaw/openclaw/releases

[12:00] STORY 2 — Project Glasswing: The Cyber Defense Coalition

Anthropic launched Project Glasswing with a coalition of Amazon, Apple,

Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA,

Palo Alto Networks and others. The centerpiece is Claude Mythos Preview —

an unreleased frontier model scoring 83.1% on CyberGym vs 66.6% for Opus

4.6. In testing it found thousands of zero-day vulnerabilities, including

a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw. Anthropic is

committing $100M in usage credits and $4M in donations to open-source

security orgs. The core thesis: offensive AI capability has outpaced human

defensive response time, so the same capability must be deployed

defensively. Worth discussing: what does "coalition" mean when Anthropic

controls the model? And is finding bugs and patching them actually better

than just not shipping vulnerable code?

→ anthropic.com/glasswing

[20:00] STORY 3 — MegaTrain: Full Precision Training of 100B+ on a Single GPU

MegaTrain enables training 100B+ parameter LLMs on a single GPU by storing

parameters and optimizer states in host (CPU) memory and treating GPUs as

transient compute engines. On a single H200 GPU with 1.5TB host memory,

it reliably trains models up to 120B parameters. It achieves 1.84x the

training throughput of DeepSpeed ZeRO-3 with CPU offloading when training

14B models, and enables 7B model training with 512k token context on a

single GH200. Practical implications: dramatically lowers the hardware

barrier for frontier-scale training, which could accelerate both

legitimate research and... everything else.

→ arxiv.org/abs/2604.05091

[27:00] STORY 4 — 178 AI Models Fingerprinted: Gemini Flash Lite Writes 78% Like Claude 3 Opus

A research project created stylometric fingerprints for 178 AI models

across lexical richness, sentence structure, punctuation habits, and

discourse markers. Nine clone clusters showed >90% cosine similarity.

Headline finding: Gemini 2.5 Flash Lite writes 78% like Claude 3 Opus but

costs 185x less. The convergence suggests frontier models are hitting

similar optimal patterns despite different architectures and training data

— or that Claude's style is just a strong attractor for RLHF. Implications

for AI detection tools, originality claims, and the economics of "good

enough" AI writing.

→ news.ycombinator.com/item?id=47690415

[32:00] STORY 5 — LLM Plays Shoot-'Em-Up on 8-bit Commander X16 via Text Summaries

A developer connected GPT-4o to an 8-bit Commander X16 emulator using

structured text summaries ("smart senses") derived from touch and EMF-

style game inputs. The LLM maintains notes between turns, develops

strategies, and discovered an exploit in the built-in AI's behavior.

Demonstrates that model reasoning can emerge from minimal structured

input — no pixels, no audio, just text summaries of game state. Fun side

note: the Commander X16 is a modern recreation of an 8-bit home computer

architecture, so it's running on actual hardware emulated in software.

→ news.ycombinator.com/item?id=47689550

[35:30] OUTRO / CLOSE

Next episode drops tomorrow. If you want a transcript, reply on Telegram.

→ Reply on Telegram to approve transcript generation.

```

Show notes: https://tobyonfitnesstech.com/podcasts/episode-26/

Show notes: https://tobyonfitnesstech.com/podcasts/episode-26/

This week’s throughline is control: who controls the runtime, who controls agent behavior during real incidents, and who controls the physical systems AI now depends on.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-25/

Show notes: https://tobyonfitnesstech.com/podcasts/episode-25/

OpenAI buys a media platform. Peter Steinberger highlights the CLI workaround culture forming around Anthropic's restrictions. Microsoft launches an open-source agent governance toolkit. Meta shows AI optimizing the machine layer underneath inference. Microsoft commits ten billion dollars to AI infrastructure in Japan on sovereignty terms. And in the United States, the data-center boom runs headfirst into an old-fashioned bottleneck: electricity. Six stories about who controls the AI stack — and whether the physical grid will let anyone finish building it.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-24/

$300 billion in one quarter. Anthropic pays $400 million for a team of nine. Google open-sources its best reasoning model. The World Economic Forum says it's time to treat AI compute like power grids and water systems. And effective today, Anthropic is changing how third-party harnesses like OpenClaw are billed — because the infrastructure era isn't just about data centers. It's about who pays for the compute. Six stories about the week infrastructure stopped being boring.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-23/

The software shipped before breakfast. OpenClaw v2026.4.1 turns background agent work into a first-class chat surface with /tasks, bundles SearXNG for private web search, and lands Voice Wake on macOS — the agent OS shift in one release. Microsoft drops three in-house foundational models on the same day and declares itself a top-three AI lab. Okta launches enterprise AI agent governance, treating every agent as a non-human identity with a kill switch. Oracle cuts thousands of jobs to fund the Stargate infrastructure bet. And the White House advocates for federal AI preemption while 45 states have already introduced 1,500+ bills — with the EU AI Act's high-risk enforcement clock ticking to August.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-22/

Three agent runtimes walked into a codebase. Only one knew what it was building toward. NOVA and ALLOY open the actual source files for OpenClaw, Claude Code, and Hermes Agent — and let the architecture tell the story. The turn cycle. The memory model. The safety system. The skills ecosystem. And the most telling detail: Hermes ships a migration tool called hermes claw migrate that imports OpenClaw skills. That tells you who set the standard.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-21/

OpenClaw stopped being a clever tool this week and started being infrastructure. NOVA and ALLOY cover five stories: the v2026.3.31 release that unified background tasks, tightened plugin security, and hardened gateway auth; OpenClaw's viral moment in China — GitHub stars past React, lobster victims, and a state crackdown; Microsoft integrating OpenClaw into Microsoft 365 for 400M enterprise users; Perplexity's always-on local Personal Computer agent; and a $297 billion Q1 2026 VC quarter where 81% went to AI. The throughline: capability without governance is a demo. Capability with governance is a product.

Show notes: https://tobyonfitnesstech.com/podcasts/episode-20/