Ken｜Startup COO & PR

Hold on. An AI just opened your desktop app, ran in the background, and submitted a Pull Request — all while you were in a meeting. OpenAI just dropped a major Codex update that goes way beyond autocomplete. We're talking: computer operation alongside you, long-running background tasks, multi-agent worktrees running in parallel, and automated review-to-diff pipelines. This isn't a coding assistant anymore. This is a junior engineer who never sleeps, never complains, and actually reads the docs. What genuinely moves me isn't the tech itself — it's the shift in *what humans get to do*. When an agent handles the repetitive scaffolding — renaming, refactoring, PR summaries, test runs — the human gets to focus on the judgment calls. The architecture decisions. The things that actually require taste. The practical entry point? Start small. Give it a docs fix or a minor bug. Make it pass CI. Then watch the feedback loop compress in real time. The teams that will win aren't the ones with the most engineers. They're the ones who learned how to *direct* agents with precision — clear task boundaries, locked environments, human checkpoints at the right moments. AI Agents aren't replacing your thinking. They're giving your thinking more room to breathe. openai.com/hr-HR/codex/ #AIAgents #OpenAICodex #AgenticCoding #FutureOfWork #DeveloperTools

Asked the AI agent to prep the quarterly IR summary. It finished in 40 seconds. Then I spent 2 hours checking if it was right. My colleague watched this and said: 'So you outsourced the writing and kept the anxiety for yourself.' I had no rebuttal. Because that is, functionally, exactly what I did. We now have a word for this at the office: productivity theater. You delegate the task. You hoard the stress. The output is faster. You are not. Standing context, spend limits, approval gates — yes, we set all of that up. The one thing we forgot to configure was a cap on how long a human can second-guess a document that a machine produced in less time than it takes to pour coffee. Somebody should file that as a feature request. support.claude.com/en/articles/13… #ClaudeCowork #AIAgent #IRComms #ProductivityTheater #AnthropicEnterprise

Background computer use is not the same as asking it to open a tab and draft something. Codex operates your desktop from behind the scenes — the way you would — but without stopping to confirm. That gap between helpful and dangerous is exactly where I got it wrong. Six weeks ago I ran a stakeholder update task for a client's IR workflow. I forgot to set the audience context. Codex defaulted to a press-facing tone. It drafted a memo that read like a newswire release, not a board document. My client James caught it at 11:02am — two minutes before his EA would have forwarded it upstream. He said: "Ken, this goes to the CFO. Not Reuters." After that, I locked three gates into every background agent deployment. First: label the audience. Board, press, IR, ops — one word, set it upfront. Second: set a tone gate. Formal, exec, external — hardcoded in the system prompt. Third: add a human checkpoint before anything touches an external send endpoint. Three comms workflows now run on Codex background mode. Zero misfires across 11 days of live use. Around 4 hours recovered per week across two clients. Maybe what I should have been thinking about from the start wasn't the output — it was who was on the other end of the send button.

Codex just filed a stakeholder update. I was in another meeting. In 2021, Codex lived inside GitHub Copilot. It completed lines. That was the whole job. This week, OpenAI shipped background computer use. Codex now sees your screen, clicks through apps, submits files. No prompt. No keyboard. It just runs. That rewrites what an AI assistant can do for anyone in comms. Point it at any workflow that involves opening apps and filling in documents. At 9:14am today, my colleague Marcus messaged: "It just sent the thing. I didn't touch anything." There's one thing I need you not to get wrong here,

Hold on. An AI just stopped itself from running a dangerous command — without anyone asking it to. That's what Claude Code's Auto mode actually does. It's not 'approve everything.' It has a permission gate built in — a safety checkpoint that quietly intercepts risky tool calls before they execute. Researchers stress-tested that gate (arXiv, April 2026) and measured false positives, missed catches, real-world edge cases. It's not perfect. But it exists. And that changes everything about how we think about autonomous AI agents. Here's what hits me: We spent years afraid that giving AI more freedom meant giving up control. Auto mode is a bet that you can have both — more autonomy AND more safety, if you design the guardrails right. That's not a small thing. That's a new philosophy of trust. For anyone building with Claude Code right now: Shift+Tab to cycle modes. Auto mode is where the real leverage is. But read the research. Know where the gate holds — and where it doesn't. The teams that will win with AI agents are the ones who understand the system deeply, not just the ones who move fastest. Autonomy without understanding is just speed toward the wrong wall. Understanding without autonomy is just theory. The sweet spot? That's what we're all building toward. arxiv.org/abs/2604.04978 #ClaudeCode #AIAgents #AutoMode #Anthropic #AIWorkflow #FutureOfWork

My team finally agreed to let an AI agent handle the first draft of our IR summary. Step 1: Set up AGENTS.md with all the rules. Forbidden phrases. Required citations. Approval gates. Tone guidelines. Step 2: Run the agent. Step 3: Spend two hours explaining to the agent why 'robust synergistic momentum' is not a financial metric. Step 4: Add 'no made-up adjectives posing as data' to AGENTS.md. Step 5: Repeat. Somewhere around revision six, my colleague looked up and said: 'We are basically parenting, but the child already knows how to code.' I had no rebuttal. The real deliverable from agentic IR workflows is not the document. It is the AGENTS.md file that slowly becomes a mirror of every mistake you forgot you made. agentsmd.io #AIAgents #IR #EnterpriseAI #ClaudeCode #AGENTS

Okay, this genuinely stopped me in my tracks. OpenAI just dropped a major Codex update — and it is not just about writing code anymore. We are talking about an AI agent that operates in the background, wakes itself up to continue long-running tasks, controls your desktop apps, and builds a memory of your work over time. Let that sink in for a second. The practical shift here is real: instead of prompting an AI and waiting, you now define a task by its outcome, its constraints, and how you will verify success — then let it run while you focus elsewhere. Background execution. Persistent context. Pull-request-ready output with diffs, test logs, and reproduction commands baked in. What moves me most is not the technology itself. It is what this unlocks for the people doing the work. A solo founder. A small comms team. A developer who has been doing the work of three people. Suddenly they have something that does not just assist — it carries. This is not about replacing human judgment. The best use cases still end with a human reading the output, approving the PR, signing off on the message. But the distance between idea and finished artifact just got dramatically shorter. If you are thinking about how to run communications, development, or research operations with a smaller team and higher standards — this architecture of outcome plus constraint plus verification is worth building into how you work right now. The agents are not coming. They are already here, asking for permissions. openai.com/index/codex-fo… #AIAgent #Codex #OpenAI #FutureOfWork #ProductivityTools #ClaudeCode #AITools

Wait — AI just reviewed the PR, ran the tests, cited the logs, and drafted the fix. All of it. In one flow. OpenAI's Codex just expanded beyond code generation into the full development lifecycle. PR review. Multi-terminal browsing. SSH into remote devboxes. A built-in browser for iterating on frontend in real time. But here's what actually hit me: The pattern they've formalized is this — Diff the PR → run the tests → surface findings with evidence (logs, citations) → propose a fix PR. That's not a tool anymore. That's a collaborator with a paper trail. I've spent years watching developers drown in review queues, context-switching between terminals, losing the thread between what broke and why. Codex is quietly handing that cognitive load back. And the deeper thing? When an AI agent leaves a structured PR — intent, verification, risk scope, rollback — human review gets faster. Not because humans do less. Because the signal is cleaner. This is what good tooling does. It doesn't replace judgment. It creates the conditions where judgment can actually land. If you're building, reviewing, or leading a dev team right now — this update is worth your 10 minutes. The loop just got tighter. openai.com/index/codex-fo… #AIAgent #Codex #OpenAI #DeveloperTools #CodeReview #SoftwareEngineering #FutureOfWork

Hold on — your AI agent just got a whole team of specialists. Gemini CLI dropped Subagents, and honestly, my brain is still catching up. Here is what changed: you drop a markdown file into .gemini/agents/, name it something like frontend-specialist.md, and from that moment on you can call @frontend-specialist directly inside your workflow. It runs in its own isolated context. No more bloated single-thread conversations where everything bleeds into everything else. Think about what that actually means. Large-scale refactors that used to spiral into chaos — because one agent was holding too much context, making too many decisions, losing coherence halfway through — can now be handled like a real team. One orchestrator. Multiple specialists. Each one focused, clean, accountable. This is not just a productivity trick. It is a fundamental shift in how we design AI-assisted work. The old model was: give one genius all the information and hope it figures it out. The new model is: build a structure, assign roles, let specialization do what specialization has always done best. And the beautiful part? The complexity lives in the architecture, not in the prompt. You stop fighting context limits. You start designing systems. If you have ever felt the frustration of an AI that starts brilliant and ends confused — this is the answer you were waiting for. developers.googleblog.com/subagents-have… #GeminiCLI #AIAgents #Subagents #AIWorkflow #FutureOfWork #AgenticAI

the one that shifts the subject from its own stack to the client's result. I ran Claude Code and Codex in parallel for three months straight. Month one: I built a 47-row comparison sheet, color-coded, dropped it in the team Slack channel. Nobody used it to make a decision. Month two: I wrote a six-page evaluation summary with performance breakdowns. The team read the opening paragraph, then asked which one I personally preferred. Three months of evaluation. Wrong question the entire time. The shortlist was never a tool question. It was always a business question, and I had been answering the one nobody was asking. Three things I changed after that: Map the exact bottleneck before you open any agent. Find the specific task where an hour disappears — not the process that "kind of feels slow." (Exception: no baseline, no story. If you cannot measure today's time, build that measurement log before anything else.) Run both agents on your actual Monday 9am task. Not a demo scenario. The real recurring one you do on autopilot. Time it. Save the raw output. Mark every error by hand. Report the delta. Only the delta. Press release drafts: 4 hours down to 40 minutes. Review cycles: 6 rounds down to 2. One number outlasts six pages of feature comparison in any room. Probably, I can't go back to evaluating agents before I've timed the work.

Your AI agent stack is on someone's procurement shortlist right now, sitting next to two other teams' stacks. Not hypothetically. OpenAI just expanded Codex to cover "almost everything" — code review, debugging, team coordination, jobs that keep running while you sleep. TechRadar called it a direct shot at Claude Code. That same week, a study landed: 7,156 real PRs, five agents compared, acceptance rates broken down by task type. The way out is simple. Stop writing "we use Codex." Write what the other side gets. Not "we run AI agents" — "your press release is ready in 40 minutes, not 4 hours." But the team that gets chosen is, →→→

Gave Codex a task that was definitely not just code. Me: 'Hey, can you reorganize my project folder, draft the status update, and flag anything that looks off?' Codex: *does all three without blinking* Me: '...so what exactly are you NOT for?' Codex: *opens browser tab* At some point the job title 'AI coding assistant' became what 'executive assistant' sounds like on a polite resume. The honest label is: ambient competence in a box. You give it a mess. It gives you back a diff and a log. No sighing. No 'that's not really my role.' Just quiet, slightly unsettling thoroughness. I told a colleague. He nodded slowly and said, 'So it's a coworker who never needs context on Monday morning.' Honestly? That's the most accurate product description I've heard. I'm keeping my job. Probably. openai.com/index/codex-fo… #Codex #AIAgent #OpenAI #VibeCoding #FutureOfWork

Okay, this one genuinely stopped me mid-scroll. Claude Cowork just hit general availability — and I don't mean 'another AI tool dropped.' I mean: role-based access controls, spending limits per team, OpenTelemetry observability, audit logs, and scheduled task automation... all inside a single workspace your comms and IR teams can actually govern. Think about what that means in practice. You're not just giving people a chatbot. You're building a system where a junior PR associate can run weekly media monitoring on a schedule, a finance team can control exactly which agents touch sensitive filings, and a CTO can see a full telemetry trace of every automated action — without anyone needing to write a line of code. That's not automation. That's institutional memory, encoded. For years, enterprise AI adoption stalled because the answer to 'who's responsible when the agent does something wrong?' was 'uh... we'll figure it out.' Cowork's GA release is the first time I've seen a platform say: here's the permission gate, here's the spend ceiling, here's the log — now go build with confidence. If you're the person at your company quietly holding together AI pilots with duct tape and hope, this is the infrastructure you've been waiting for. The future of communications operations isn't faster drafting. It's structured, auditable, human-in-the-loop intelligence at scale. insights.marvin-42.com/articles/claud… #ClaudeCowork #AIAgent #Anthropic #EnterpriseAI #PRtech #IRtech #AIGovernance #CommunicationsStrategy

My colleague just watched me set up an AI agent to auto-draft a competitive analysis, cross-reference three tabs of public filings, and format a summary report. He stared for a second, then said: 'So you outsourced your thinking.' I said no, I outsourced the part that felt like dragging furniture across a wet floor. He thought about it. '...So you outsourced your thinking.' Look, there is a difference between thinking and furniture-dragging. I have decided that difference is billable hours. tomsguide.com/ai/google-just… #AIAgent #Gemini #Claude #Productivity

Asked an AI agent to help prep our IR deck. It finished in four minutes. Then it asked if it should also 'monitor the approval chain and flag human bottlenecks.' I stared at the screen for a while. Because the bottleneck it was referring to... was me. The permission gate was working exactly as designed. It was just waiting on the slowest node in the workflow. Turns out the hardest part of automating PR/IR isn't the drafting or the formatting or even the compliance check. It's the moment the system politely surfaces the fact that you are the manual step. I approved the output. I did not approve the observation. openai.com/index/introduc… #AIAgent #Codex #PR #IR #EnterpriseAI

Told the AI agent to handle a long-running task while I grabbed coffee. Came back. It had filed a PR, written the test suite, updated the docs, and left a note in the commit message that said: 'Awaiting human review.' I stood there holding my mug. It wasn't waiting for my expertise. It was waiting for my signature. There's a difference. A quietly devastating one. I approved it. Obviously. The tests passed. But something about being the 'human review checkpoint' between two AI handoffs makes you feel less like a senior engineer and more like a notary public who learned to code. openai.com/index/introduc… #AIAgents #CodexAI #ClaudeCode #DevLife #HumanInTheLoop

Asked our AI agent to help draft the Q1 investor update. It came back in 4 minutes with a clean structure, three data visualizations, and a risk disclosure section I definitely would have forgotten. I stared at it for a moment and thought: this thing just did in 4 minutes what used to ruin my Tuesday. So I did what any reasonable comms professional does. I changed two words, moved a comma, and sent it to legal feeling like I had earned my afternoon. There is a name for what I experienced: Supervised Authorship. It is when the AI writes and you watch very carefully to make sure it matches your vision, which is definitely real and was always there. help.openai.com/en/articles/11… #AIAgent #Codex #PRandIR #SupervisedAuthorship #ClaudeCode

Hold on. Let me make sure I'm reading this right. OpenAI Codex just became an agent that clicks, types, remembers, schedules, and talks to Jira, Slack, and Notion — all while you sleep. Not a chatbot. Not a code autocomplete. A teammate that runs the whole loop: change → verify → report → done. What blows my mind isn't the tech itself. It's what this means for how we work. For years, 'automation' meant 'write a script and pray it doesn't break at 2am.' Now? You describe the workflow in plain language, set an approval gate for anything sensitive, and let the agent handle the rest. The part that genuinely moves me: Codex remembers. It builds skills from repeated tasks and reuses them. That's not just efficiency — that's institutional knowledge that doesn't walk out the door. For anyone in communications, PR, or ops: Imagine your weekly briefing pack drafted automatically, with sources cited, calculations shown, and nothing sent externally without a human sign-off. That's not science fiction as of April 2026. That's Tuesday. The question worth sitting with: If an agent can now handle the mechanical repetition of your job, what becomes possible when your full human attention goes somewhere only you can take it? That answer is worth chasing. agenticbrew.ai/news/8f66b820-… #AIAgent #OpenAICodex #FutureOfWork #Automation #PRtech #CommunicationsStrategy

Wait — your AI coding agent can now work for DAYS without you babysitting it. OpenAI just dropped a major Codex update, and the headline isn't just 'better code generation.' It's this: background execution, PC app control, and task scheduling across days or even weeks. Codex can now wake itself up, pick up where it left off, and keep building — while you sleep. But here's what actually gives me chills. The real craft isn't in the model. It's in how YOU design the handoff. Break the mission into milestones. Define the wake conditions. Lock down the output folder. Draw a hard line around what the agent can execute, delete, or send — without asking you first. That's not just software engineering. That's a new kind of trust architecture between humans and machines. We're not at 'AI does everything.' We're at 'AI carries the weight, humans hold the wheel.' And honestly? That's the more exciting story. The builders who'll win this era aren't the ones who hand everything to the agent. They're the ones who design the guardrails so beautifully that the agent runs fast AND safe. If you've been waiting for the moment where AI agents stop being a demo and start being a genuine work partner — that moment is uncomfortably close. The question isn't whether to use it. The question is: have you designed your rules of engagement yet? openai.com/index/introduc… #AIAgent #Codex #OpenAI #FutureOfWork #ClaudeCode #BuildersOfAI #HumanInTheLoop

Asked the AI agent to handle a PR draft while I grabbed coffee. Came back to find it had also reorganized the repo, flagged three security gaps, opened two PRs, and left a polite comment asking if I wanted it to schedule the press release too. I said yes. It is now Wednesday. Here is what I have learned: there is a meaningful difference between 'delegating a task' and 'handing someone the keys to your entire workflow and walking away for eight minutes.' I call the second one 'aggressive coffee.' The agent is fine. The PRs are clean. The press release is honestly better than mine. I am choosing to frame this as a win rather than think too hard about what it implies. openai.com/index/codex-fo… #AIAgents #CodexAI #ClaudeCode #Automation #AgentMode