DevDay 2026 highlights
This was my second DevDay.1 Building through iterations of APIs like Chat Completions and Assistants (RIP) + plugins (GPT store, etc.) and variation across labs has underscored the importance of staying provider agnostic.2
I’ve been thinking a bit about how general agents will play with more specialized ones.
To make the experience work for consumer agents at acceptable economics, a lot has to happen under the hood. Big companies have an advantage here today because so much is helped by scaled compute / trust in security infra / distribution. For companies paying API prices, running complex consumer workflows can still cost a handful of dollars or more a day per user, even if they’re partially routed to cheaper models, cached, batched etc. Yes, token costs continue falling, but the frontier also keeps moving. Charging an additional $50-100+ monthly for non-work-related uses is a hard sell.
So Sign In with ChatGPT is great for anyone building specialist agents; it couldn’t come at a better time for me. If a general AI subscription ends up as standard as a phone bill, someone who’d balk at even another $20 or $40 a month can still try a service with usage included in the one they already pay for. Whatever giants also enable this will subsidize us for a while.
But the other, more underrated part of DevDay for me was support for MCP Events in the latest rev of Plugins. Sign In with ChatGPT helps with usage costs; plugins help with distribution. MCP Events now lets developers expose specialist agents’ events for ChatGPT to act on.
For an agent to be useful, it has to remember + be proactive (& eventually, learn). Proactivity so far has been either scheduled tasks / crons or event-based. Basically, some personal agents (like OpenClaw) wake themselves up on a pre-determined cadence, e.g. every half-hour, to see if there’s something they ought to do and/or touch base with you on.3
With event-driven proactivity, instead of waking up on a schedule to check for new calendar invites, the system gets notified when one comes in. This is better when the trigger is something happening, but doesn’t cover cases where the important signal is that something hasn’t happened yet.4 Both are necessary, so specialized agents may still benefit from maintaining their own awareness of time, expectations and scheduled wakeups.
MCP Events give Dots (+ Muse, Instinct, etc. when they hopefully follow suit) - or whoever wins general agents - a standard way to be woken up by specialized agents.
For example, let’s say a specialized money-saving agent called Frugal (‘Frug’, for short) is adept at disputing all kinds of personal finance nightmares: rising car insurance, expiring discounts on internet bills (screw you Xfinity), asinine medical bills, price gouging on storage units, etc. Frug’s developers have optimized this to work way better than native personal finance capabilities in ChatGPT for various reasons.
If Frug builds a plugin, it can now ping your Dot when it needs to inform you, for example, that your insurer’s renewal quote came in 28% higher, and that it has some ideas to respectfully fight this. Specialists like Frug can compete on capability and a more bespoke experience without forcing us to hold their hands separately.
Specialized agents ought to expose functionality in multiple general aggregators, like companies ship iOS & Android apps. Negotiating power with their suppliers on one end and general agents on the other will depend in part on how rare the context accumulated through use of the specialist is and how proprietary their merchant supply is.
A few more notes, though much ink has been spilled on these already:
- Some argue that agent-mediated commerce could leave merchants with lower margins per sale but increase volume compared to ads. I think this is likelier in discretionary categories with higher natural ceilings, where demand can grow. Think of certain aspects of health, like Korean skincare or recurring aesthetic treatments. Specialized agents should aggregate supply and trade merchants volume for favorable terms, so users prefer transacting through them.
- Yes, some things like fighting price hikes or getting quotes from contractors will be useful to automate. But we enjoy shopping and putting trips together for the sake of it, so specialists should build to enhance the fun of the experience itself as well as what they automate.
- IMO people prefer privacy, discretion and ownership of their identity / context IF it doesn’t come at the cost of convenience. I don’t think self-hosting the general agent is out of the question if it’s made easy… which would make Apple the obvious company to make that possible on personal devices, if they didn’t have such inconsistent software lately.
We’ll see what happens before the next DevDay!5
-JJB
Footnotes
-
Highlight for me last year was storyboarding workflows using Sora - since discontinued - so what I took away this time around might similarly change.
This year’s Sora might be voice. Voice is interesting because it has improved so much with GPT-Live, but one flat or misread moment in a live conversation (of which there were many during demos) is enough to deflate the energy in the room / give the user secondhand embarrassment.
Plus, as someone who does it all day, I literally hate talking to AI. It makes my blood boil. Opus 5.5 and Astra improve on this, but I think many of our interactions should be one way dictation with visual + non-verbal responses. In a tactile / more physical environment. ↩
-
Especially in areas like healthcare, devs need control over harness specifics like compaction, so I don’t love more out-of-the-box things like the Agents API. ↩
-
Note this is open-ended: it looks at memory and the user’s state and decides “what needs following up”. This is a generalized version of specific tasks you can schedule (like what Claude and ChatGPT do if you ask them to remind you, for instance, what meetings you have each morning.) ↩
-
E.g. “we should have seen that invite come in by now, why has it not?” wouldn’t happen if the only trigger is the invite arriving, because the system doesn’t wake to find this out until the invite comes in, which it hasn’t. ↩
-
On a general note, well-intentioned shibboleths aside, providers will prob be helpful to partners until they’re not. It’s been sort of hard to figure out where the big players fall on this spectrum today, much less where they will in the future. As things have played out, though, it’s gotten easier to be more independent without sacrificing product quality.
The cost of being prudent has dropped too. OpenAI is reporting pretty looking charts for plugin growth. Part of this is likely because it costs less to ship. From my standpoint, last year I would not have the bandwidth to ship a plugin. Sure, plugins have continued evolving as an offering, but it’s also easier to tell Astra or Opus 5.5 to build, verify, and maintain this. Strategic optionality has become an order of magnitude cheaper.
I think tokens will increasingly go into shoring up resilience in a company’s direction. This means building and maintaining more integrations across surfaces and reallocating effort over time. This turns software development into something like portfolio management, where, given strategic shifts in the market, we reallocate compute to double down on winning theses as the fog clears (or continuously shifts). I’m obviously not the first to say something like this, just saying this is what it feels like in the weeds. ↩