AI Expiry Week: GPT-5.5 Exposes the AI Lifecycle Problem

Lately, we’ve been watching enterprise software age in dog years. The retirement notices are getting hard to keep up with.

Microsoft Project Online went on September 30. Project LTSC 2021 loses support on October 13, although it'll still run without security fixes or technical support. A day later, GPT-5.5 is scheduled to leave ChatGPT, ChatGPT Work and Codex.

It's an odd week to be promising everyone a new digital coworker, especially now that Workday Rising is kicking things into high gear, with Aneel Bhusri and Gerrit Kazmaier discussing agents “doing real work.” Asana's Work Innovation Summit lands on GPT-5.5's retirement day.

OpenAI spent September expanding what its agents could handle. Team Tasks can now manage recurring project updates, keeping track of progress, blockers, and upcoming work. It's another development in the workspace-agent story we've been following. Of course, retiring the model behind a workflow won't make the reporting deadline disappear. Someone still has to deliver Monday's update.

Joshua Goldfein at Mercury Digital has a sensible approach. In an October 7 account of his team's content workflow, he explained how they keep instructions, source records, and checks separate from the GPT-5.5 and Claude models doing the work.

“Nothing in that chain depends on which model wrote which part.”

That means the team can change models without rebuilding the workflow around them. It's a detail I'd be asking vendors about while they're showing off their latest agents.

Apparently, project managers have another deadline to track: when their software workers are due for retirement.

TL;DR:

• GPT-5.5 has given project managers another deadline to worry about. October 14 is retirement day, and OpenAI wants customers checking scheduled tasks, custom agents, and scripts before the switch. Anyone relying on the model for weekly project updates should probably test the replacement. Finding out on Friday that Wednesday's model change messed up the reporting isn't much fun.

• Atlassian is giving AI agents a bigger share of the backlog. At Team ’26 Europe, Jira showed off bulk agent assignments across up to 50 work items, alongside new capacity and budget-planning capabilities. Handy, but fifty assigned tasks aren't fifty finished tasks. Someone still needs to know what's happening with them, especially when an agent gets stuck.

• Workday has introduced Sana as the new Workday home, plus Agent Harness and Custom Agents are generally available, and the first autopilot recruiting capability is in early access. The real test is whether approvals and handbacks survive when work crosses applications.

• Zoomtopia has a more interesting problem to solve than meeting summaries. Zoom can already turn conversations into tasks and trigger workflows across connected tools. Next week's event should tell us how well that holds together once work reaches Jira or another project system. A task without the original decision, a clear owner, or enough context to act on it is another thing for an already busy project manager to chase.

What Happens When GPT-5.5 Retires on October 14?

October 14 comes with a rather frustrating project-management job: finding every recurring task that still expects GPT-5.5 to turn up for work. OpenAI is retiring it from ChatGPT, ChatGPT Work, and ChatGPT-signed-in Codex. API access is unaffected by this particular change. Its instructions are wonderfully specific:

“Update workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks, and scripts.”

That's quite a scavenger hunt for whoever owns the project plan.

At DevDay, OpenAI demonstrated Team Tasks assembling weekly updates from project documents, tracking progress, blockers, and next steps. Reuters puts weekly Codex and ChatGPT Work usage above 35 million. Nobody knows how many scheduled reports depend on one model, which is pretty much the problem.

Changing the model name might be the easy part. In OpenAI's developer community, Corawz described testing the same work after switching:

“My code runs fine on 5.6; what does not survive the swap is a week of work fitting inside a week of allowance.”

One developer's bill isn't everyone's, but it's reason enough to test costs alongside accuracy and permissions.

A September preprint examining 22,555 migration commits estimated that 82% of migrations away from retired models happened after shutdown. That's open-source evidence rather than a survey of PMOs, but hardly reassuring.

So audit those dependencies and rerun the actual tasks. October 23 brings a separate API retirement batch, including older GPT-3.5 Turbo, GPT-4, and o-series models, so it’s worth preparing for that too. Apparently, project managers need a model-expiry column these days.

What Happens to Gemini Gems as Google Moves to Skills?

Google's October 8 announcement came with an ambitious promise: a universal Gemini agent that can get on with work across Workspace and connected apps, complete with model choice, access controls and spending oversight. It's in private preview for now, so I'd like to see what happens when an agent passes a project task between apps and someone asks who's responsible for it.

Meanwhile, Google is also rearranging the assistants people already use. Gems, its customizable helpers for repeatable tasks, are gradually making way for Skills. If you've built your project routines around Gems, there's some calendar admin coming.

The timing's already changed. On October 13, Google said Skills were starting to roll out in the Gemini app, but delayed the Workspace release until further notice. The old October 5 and October 19 Workspace milestones are out. Gems still move into Gemini's Settings panel on November 17, and Business and Enterprise users can keep creating and using them until no sooner than March 1, 2027

There's still homework, though. Existing Gems will eventually become draft Skills, but Skills created in the Gemini app won't automatically appear in Workspace. You'll need to recreate them there. Existing Workspace Studio flows using "Ask a Gem" also keep working, although Google has stopped people building new ones.

The migration timetable gives teams some breathing room. Skills also use the portable Markdown-based SKILL.md format, which should make moving instructions between tools easier.

I'd still check those recurring project workflows before celebrating. Nobody wants to discover their carefully configured assistant has taken the weekly status report with it.

Which Project Platform Shipped the Week’s Most Useful Update?

OpenAI wasn't quite finished after DevDay. On October 9, it gave Dots some more responsibilities: starting work in Codex, following existing threads and reviewing or editing Scheduled Tasks in ChatGPT Work.

That's a fairly interesting development for anyone managing recurring project work. The October 9 update also arrived just days before GPT-5.5's retirement. We've already covered OpenAI's growing task-management ambitions, but this makes the question of who maintains those automations harder to ignore.

Still, Atlassian might have the more convincing project-management story this week. At Team ’26 Europe, Atlassian made a decent case for giving agents a proper place on the project team.

Jira can already assign work to agents, delegate tasks in bulk, and track their activity. New capacity views and project-budgeting signals give managers more to work with, while the Agentic Multiplayer Protocol offers shared context and clearer permissions. Agent accounts and a non-human identity inventory are also coming.

Automation Consultants had one of the more convincing examples. Its PMO uses 12 agents across 11 client engagements and reports a 25% gain in delivery capacity, alongside three times the throughput. I liked the decision to keep agent state in files instead of conversation histories, making model changes easier. There's oversight, too. A supervising agent reviews findings, and people approve changes before they reach client systems.

Jose Ramirez, of Kaplan North America, made a strong claim too, saying its agents can:

“summarize a month’s worth of work across five different platforms in 10 seconds.”

Microsoft's September Planner update points in a similar direction, with Planner Agent creating tasks, assigning owners, and working across larger plans. It even has a model selector.

That's encouraging, since project managers have enough trouble keeping track of human colleagues without losing the history whenever an AI model changes.

What Did Workday Announce at Rising 2026?

Tuesday's keynote happily went beyond the somewhat vague promise of agents ‘doing real work’ with some actual answers about where those agents get their instructions, permissions, and approval trails. Before the event, Gerrit Kazmaier said: