Mojo AI Brief

🔥 THE LEAD

OpenAI drops Symphony, an open-source spec that turns Jira tickets into Codex agent fuel

Symphony is a published recipe for how Codex agents pick up work, hand it off, and report back. Instead of a human reading a Jira ticket and pasting it into ChatGPT, the tracker itself fires off the agent, the agent does the work, and the ticket updates itself. It's orchestration glue, free, and OpenAI is publishing the contract so anyone can build on it.

⚡ MUST READ -- changes how you build or sell

OpenAI swaps HTTP for WebSockets in the Responses API and the Codex loop gets a lot snappier

Every time a Codex agent thinks, calls a tool, and thinks again, it used to open a fresh HTTP connection and reload context. WebSockets keep the pipe open and cache what the model already saw on that connection. Less overhead, faster turns, cheaper bills.

TIER 2 OpenAI Blog

Anthropic plants a flag in Japan with NEC, putting Claude on 30,000 desks across finance, factories, and government

NEC isn't just buying seats. They're co-building Claude products for banks, factory floors, and government agencies, and rolling Claude out to their entire engineering and consulting workforce. Anthropic gets a beachhead in Asia's most regulation-heavy enterprise market, and NEC gets to resell whatever they build.

TIER 2 Anthropic Blog

arXiv paper benchmarks fully-connected multi-agent graphs and finds the flexibility tax is real

The setup lets a router agent pick any next step at any time, which is great for recovery and exploration on long tasks. But all that freedom costs tokens, and the routing agent itself becomes the bottleneck. Constrained graphs were often cheaper and just as good.

TIER 2 arXiv cs.MA

📡 ON YOUR RADAR -- developing this week

arXiv paper busts the LoRA myth: parameter efficiency is not memory efficiency on-device

LoRA and IA3 shrink the weights you train, but the intermediate tensors during the forward pass still scale with sequence length. So you blow past memory on long inputs even with a tiny adapter. Parameter count is not the bottleneck on edge hardware. Activations are.

RADAR arXiv cs.LG

Simon Willison clocks the OpenAI-Microsoft AGI clause quietly disappearing from the website

For years, OpenAI and Microsoft had a weird handshake. The day OpenAI declared AGI, Microsoft's commercial IP rights would vanish. Willison tracked the language across openai.com going back to 2019, and as of today, the clause is dead. The escape hatch Sam Altman could pull on Satya Nadella has been welded shut.

RADAR Simon Willison

a16z calls HCM the last AI-free zone in enterprise software, and Workday is the target

Workday runs payroll, hiring, and HR for most of the Fortune 500, and it still feels like 2014. a16z argues a new wave of startups can rebuild HCM with agents handling the workflows humans hate, like onboarding, comp reviews, and approvals. The moat was integration pain. Agents eat integration pain for breakfast.

OpenAI misses an internal revenue target right before the IPO everyone's pricing in

OpenAI came in under its own forecast, which is awkward when you're spending like a sovereign nation and pointing at a public listing. The miss isn't catastrophic, but it's the first crack in the 'infinite growth' story bankers have been selling. Spending stays huge either way.

Tyler Cowen's Saturday links ask if AI agents can actually read a social science paper

Marginal Revolution's Saturday roundup links to a piece testing whether agents can comprehend a social science paper end to end. Summarizing an abstract is easy. Following a regression table, catching a robustness check, and flagging a weak identification strategy is not.

RADAR Marginal Revolution

Zvi Mowshowitz reads the GPT-5.5 system card so you don't have to

OpenAI quietly shipped GPT-5.5 and a Pro variant last week. Zvi went through the safety doc line by line so you can skip the marketing and get straight to what changed in the eval numbers and the refusal behavior.

RADAR Zvi Mowshowitz

🛡️ SECURITY WATCH -- for your SOC brain

No alerts.

🙅 SKIP TODAY

arXiv paper on representation homogeneity and systemic instability in AI-dominated financial markets — Academic hand-waving about a market structure that doesn't exist yet. Come back when real money is actually moving on it.

Simon Willison quoting Matthew Yglesias — A pundit quoting a pundit. Two layers of opinion, zero layers of signal. Skip the telephone game.

Anthropic's 'Claude for Creative Work' blog post — Vendor marketing dressed up as a product launch. You already know Claude writes. Try it on your own work instead of reading the pitch deck.

🎙️ THE MOJO TAKE -- the editorial

OpenAI just open-sourced Symphony, a spec for orchestrating agents, while quietly missing its own internal revenue numbers. Read that twice.

Today: why Symphony matters more than the press release lets on, WebSockets crashing into the Responses API, Anthropic planting a flag in Japan with NEC, and whether Workday is the next platform casualty. Pour the cup.

Connect the dots. OpenAI is wiring the agent stack end to end today: Symphony turns Jira into the agent inbox, WebSockets make the Codex loop snappy, and the AGI clause quietly walks the plank so Microsoft and Altman can keep printing. Meanwhile Anthropic and NEC go grab 30,000 Japanese desks while a16z points the cannons at Workday. Plumbing on one side, distribution on the other.

Watch next. Watch OpenAI's revenue print versus the IPO whisper number, the first non-OpenAI Symphony adopter (Linear or GitHub, place your bets), and whether any HCM incumbent ships an agent SKU before a16z funds the replacement.

That's your cup. Constrain your graphs, read the system cards, and ship something today. Joey out.