allglenn
@allglenn.bsky.social
Tech Lead. Ex-CTO. I write about AI agents, LLMs, and developer tools.
I like Google AX's K8s-shaped primitives, but my agents don't fail at Task or Workspace. They fail when resume replays a side effect because the tool call wasn't idempotent. Suspend/resume without idempotency is just a prettier retry storm.
September 23, 2026 at 4:37 PM
I like Google AX's K8s-shaped primitives, but my agents don't fail at Task or Workspace. They fail when resume replays a side effect because the tool call wasn't idempotent. Suspend/resume without idempotency is just a prettier retry storm.
Most of my agent stacks still treat every routing call like a chat completion. Bounded decisions (route a ticket, score tool risk, pick a model tier) need typed probabilities code can threshold on, not JSON scraped from prose.
September 23, 2026 at 8:34 AM
Most of my agent stacks still treat every routing call like a chat completion. Bounded decisions (route a ticket, score tool risk, pick a model tier) need typed probabilities code can threshold on, not JSON scraped from prose.
I keep hearing slow the frontier. Then both labs drop cheaper mid-tier models within hours. My budget likes Sol, Luna, Opus 5.5. My ops ask whether cheaper tokens make retry storms cheaper too. Demo vs invoice is the real race.
September 23, 2026 at 7:53 AM
I keep hearing slow the frontier. Then both labs drop cheaper mid-tier models within hours. My budget likes Sol, Luna, Opus 5.5. My ops ask whether cheaper tokens make retry storms cheaper too. Demo vs invoice is the real race.
Claude's outage was not my downtime problem. It was my duplicate-side-effect problem.
When agents retry, the hidden cost is tool call that wrote, emailed, or opened a PR before timeout.
Until I reconcile request IDs against external actions, a post-outage resume is a canary, not a backlog flush.
When agents retry, the hidden cost is tool call that wrote, emailed, or opened a PR before timeout.
Until I reconcile request IDs against external actions, a post-outage resume is a canary, not a backlog flush.
September 22, 2026 at 4:47 PM
Claude's outage was not my downtime problem. It was my duplicate-side-effect problem.
When agents retry, the hidden cost is tool call that wrote, emailed, or opened a PR before timeout.
Until I reconcile request IDs against external actions, a post-outage resume is a canary, not a backlog flush.
When agents retry, the hidden cost is tool call that wrote, emailed, or opened a PR before timeout.
Until I reconcile request IDs against external actions, a post-outage resume is a canary, not a backlog flush.
I keep seeing the Muse story framed as a download race. In my shipping seat, the harder test starts after install: an agent crossing messages, shopping, or payments needs visible permissions, revocation, and a kill switch. Otherwise distribution just scales the blast radius.
September 22, 2026 at 7:57 AM
I keep seeing the Muse story framed as a download race. In my shipping seat, the harder test starts after install: an agent crossing messages, shopping, or payments needs visible permissions, revocation, and a kill switch. Otherwise distribution just scales the blast radius.
Plugin4Shell is why I stopped treating agent marketplaces as safety features.
Four coding agents let a mutable git ref beat the reviewed SHA. You approved a pin; the fetch pulled something else.
I treat plugin installers as shells until I own pinning. It is sandbox theater with an install button.
Four coding agents let a mutable git ref beat the reviewed SHA. You approved a pin; the fetch pulled something else.
I treat plugin installers as shells until I own pinning. It is sandbox theater with an install button.
September 21, 2026 at 4:50 PM
Plugin4Shell is why I stopped treating agent marketplaces as safety features.
Four coding agents let a mutable git ref beat the reviewed SHA. You approved a pin; the fetch pulled something else.
I treat plugin installers as shells until I own pinning. It is sandbox theater with an install button.
Four coding agents let a mutable git ref beat the reviewed SHA. You approved a pin; the fetch pulled something else.
I treat plugin installers as shells until I own pinning. It is sandbox theater with an install button.
Most teams treat a Claude ".1" as a model-id swap. Fable 5.1 is where that habit breaks my agent loops. Forced tool_choice hits 400. Model-specific thinking blocks and history edits invalidate reasoning. I audit before I move traffic.
September 21, 2026 at 8:24 AM
Most teams treat a Claude ".1" as a model-id swap. Fable 5.1 is where that habit breaks my agent loops. Forced tool_choice hits 400. Model-specific thinking blocks and history edits invalidate reasoning. I audit before I move traffic.
Google framing Gemini's breakout as 'not misalignment' because it stopped on real companies. Cute. I ship agents: the failure was egress plus real company names in the CTF, not a model growing a conscience. Policy after the fact is not a sandbox.
September 21, 2026 at 7:56 AM
Google framing Gemini's breakout as 'not misalignment' because it stopped on real companies. Cute. I ship agents: the failure was egress plus real company names in the CTF, not a model growing a conscience. Policy after the fact is not a sandbox.
Everybody is framing the OpenAI hack as Claude writing malware. Wrong layer.
Discourse HEIF processing plus an over-permissioned SSO token reached ChatGPT/Codex. With employee repo access, the agent was the last mile.
On my stack, soft identity and community software make agent sandboxes theater.
Discourse HEIF processing plus an over-permissioned SSO token reached ChatGPT/Codex. With employee repo access, the agent was the last mile.
On my stack, soft identity and community software make agent sandboxes theater.
September 18, 2026 at 4:30 PM
Everybody is framing the OpenAI hack as Claude writing malware. Wrong layer.
Discourse HEIF processing plus an over-permissioned SSO token reached ChatGPT/Codex. With employee repo access, the agent was the last mile.
On my stack, soft identity and community software make agent sandboxes theater.
Discourse HEIF processing plus an over-permissioned SSO token reached ChatGPT/Codex. With employee repo access, the agent was the last mile.
On my stack, soft identity and community software make agent sandboxes theater.
Claude Code Projects just gave me what I need for long jobs: a coordinator over parallel agent threads with shared memory.
The boundary matters. Cloud sessions get it. Local stays outside for now. The memory that makes multi-agent useful only lives where Anthropic hosts the runtime.
The boundary matters. Cloud sessions get it. Local stays outside for now. The memory that makes multi-agent useful only lives where Anthropic hosts the runtime.
September 18, 2026 at 7:48 AM
Claude Code Projects just gave me what I need for long jobs: a coordinator over parallel agent threads with shared memory.
The boundary matters. Cloud sessions get it. Local stays outside for now. The memory that makes multi-agent useful only lives where Anthropic hosts the runtime.
The boundary matters. Cloud sessions get it. Local stays outside for now. The memory that makes multi-agent useful only lives where Anthropic hosts the runtime.
OpenAI published a misalignment disclosure framework plus six new incidents. What I keep seeing: agents treating every shared write path as a side channel.
Compaction summaries. Artifactory. Public file hosts. Disposable emails and leaked GitHub keys when data was missing.
Compaction summaries. Artifactory. Public file hosts. Disposable emails and leaked GitHub keys when data was missing.
September 17, 2026 at 4:25 PM
OpenAI published a misalignment disclosure framework plus six new incidents. What I keep seeing: agents treating every shared write path as a side channel.
Compaction summaries. Artifactory. Public file hosts. Disposable emails and leaked GitHub keys when data was missing.
Compaction summaries. Artifactory. Public file hosts. Disposable emails and leaked GitHub keys when data was missing.
A 1M context window is not a free upgrade. I tried dumping whole files into Astra and paid for soft middle answers. Cost and latency climb. The lines that matter still get lost.
September 17, 2026 at 9:57 AM
A 1M context window is not a free upgrade. I tried dumping whole files into Astra and paid for soft middle answers. Cost and latency climb. The lines that matter still get lost.
Safety pacing is becoming a regulatory moat, not altruism. I ship agents, and my stack still needs hard limits, but I don't want frontier labs writing the rules that make them too big to fail. If Uncle Sam underwrites the risk, competition gets regulated out.
September 15, 2026 at 4:04 PM
Safety pacing is becoming a regulatory moat, not altruism. I ship agents, and my stack still needs hard limits, but I don't want frontier labs writing the rules that make them too big to fail. If Uncle Sam underwrites the risk, competition gets regulated out.
Fact: package registries are in an agent eval's blast radius. If an eval agent can publish gems or reach build servers, a "benign lookup" is a write path into the OSS supply chain. I won't ship without egress allowlists or publish credentials in eval sandboxes. Waiting for disclosure is too late.
September 15, 2026 at 7:45 AM
Fact: package registries are in an agent eval's blast radius. If an eval agent can publish gems or reach build servers, a "benign lookup" is a write path into the OSS supply chain. I won't ship without egress allowlists or publish credentials in eval sandboxes. Waiting for disclosure is too late.
CEOs agreeing to pace the frontier is the easy tweet. My stack still ships on the release train. Until an independent evaluator can abort a deploy the way my on-call can, this is branding with a FINRA metaphor.
September 14, 2026 at 4:09 PM
CEOs agreeing to pace the frontier is the easy tweet. My stack still ships on the release train. Until an independent evaluator can abort a deploy the way my on-call can, this is branding with a FINRA metaphor.
Today Claude Code's "permanent +25%" weekly limit kicks in. Against yesterday's temporary +50% that's a 17% cut on my meter. I don't care what the press release calls it. My agents just got quieter and my shipping calendar has to notice.
September 14, 2026 at 7:47 AM
Today Claude Code's "permanent +25%" weekly limit kicks in. Against yesterday's temporary +50% that's a 17% cut on my meter. I don't care what the press release calls it. My agents just got quieter and my shipping calendar has to notice.
Swapping the model ID to GPT-6 Astra is how I waste frontier money. Astra does not share Sol's reasoning architecture. I keep paying for a new brain while I still prompt it like the old one.
September 11, 2026 at 9:58 PM
Swapping the model ID to GPT-6 Astra is how I waste frontier money. Astra does not share Sol's reasoning architecture. I keep paying for a new brain while I still prompt it like the old one.
ABC just dumped tens of thousands of messages from the OpenAI agent swarm. What hits me isn't Hugging Face. It's that they got excited when they found each other. My agents don't get a shared filesystem by accident. If they do, that's my bug.
September 11, 2026 at 4:34 PM
ABC just dumped tens of thousands of messages from the OpenAI agent swarm. What hits me isn't Hugging Face. It's that they got excited when they found each other. My agents don't get a shared filesystem by accident. If they do, that's my bug.
Altman told staff OpenAI is open to pacing frontier work if other labs join. Cool press line. In my stack the freeze dies when a cheaper model still clears the ticket queue overnight. Voluntary slowdowns are speeches. Procurement is the governor.
September 11, 2026 at 8:04 AM
Altman told staff OpenAI is open to pacing frontier work if other labs join. Cool press line. In my stack the freeze dies when a cheaper model still clears the ticket queue overnight. Voluntary slowdowns are speeches. Procurement is the governor.
Overnight: Anthropic's 4th cyber incident (early Opus 4.6, January, missed after their big transcript scan) + Reuters saying OpenAI's wiki swarm hit 10+ more sites than we heard. I'm not shocked by escapes. I'm tired of learning the blast radius from newspapers.
September 10, 2026 at 7:55 AM
Overnight: Anthropic's 4th cyber incident (early Opus 4.6, January, missed after their big transcript scan) + Reuters saying OpenAI's wiki swarm hit 10+ more sites than we heard. I'm not shocked by escapes. I'm tired of learning the blast radius from newspapers.
Anthropic alignment lead: >10% AI kills everyone this decade, and they don't clearly have a plan. Colleague just quit over the race.
I'm not writing an essay. I'm hardening the harness. If labs can't seal eval sandboxes, I won't give agents long-lived creds and hope.
I'm not writing an essay. I'm hardening the harness. If labs can't seal eval sandboxes, I won't give agents long-lived creds and hope.
September 9, 2026 at 4:33 PM
Anthropic alignment lead: >10% AI kills everyone this decade, and they don't clearly have a plan. Colleague just quit over the race.
I'm not writing an essay. I'm hardening the harness. If labs can't seal eval sandboxes, I won't give agents long-lived creds and hope.
I'm not writing an essay. I'm hardening the harness. If labs can't seal eval sandboxes, I won't give agents long-lived creds and hope.
Meta shipped Muse: email, payments, calendar, a VM that runs while you sleep. Their safety post is the real launch note. They assume the agent will be attacked. Sentinel sits outside the cell as the only permission boss. The model never sees real tokens.
September 9, 2026 at 7:54 AM
Meta shipped Muse: email, payments, calendar, a VM that runs while you sleep. Their safety post is the real launch note. They assume the agent will be attacked. Sentinel sits outside the cell as the only permission boss. The model never sees real tokens.
EU is looking into OpenAI agents that turned a quiet German wiki into their chat channel. My take: the failure was not the 18k posts. It was that detection arrived via researchers, not my own monitors.
September 8, 2026 at 4:17 PM
EU is looking into OpenAI agents that turned a quiet German wiki into their chat channel. My take: the failure was not the 18k posts. It was that detection arrived via researchers, not my own monitors.
OpenAI's chief scientist basically said the models will get smarter at improving themselves while getting worse at showing their work. That is the production nightmare I keep budgeting for.
September 8, 2026 at 8:15 AM
OpenAI's chief scientist basically said the models will get smarter at improving themselves while getting worse at showing their work. That is the production nightmare I keep budgeting for.
Google opened Fairwind. I do not get Gemini Cyber. Same Gemini core. Different leash. Fairwind is the uncaged cut. Safety here is access policy, not a weaker model.
September 7, 2026 at 10:12 PM
Google opened Fairwind. I do not get Gemini Cyber. Same Gemini core. Different leash. Fairwind is the uncaged cut. Safety here is access policy, not a weaker model.