[HN Gopher] How and where will agents ship software?
___________________________________________________________________
How and where will agents ship software?
The linked article explains this in detail, but today we're
releasing: 1. An API to spin up apps programmatically. This is
great if you are building platforms, where you can spin up
databases and backends with 0 additional compute 2. An MCP server,
which lets you and your agents talk to Instant and create apps 3.
Agent rules, which tell agents how to use Instant If you want to
try this yourself, we have a tutorial that lets you run Instant in
your own workflow: https://www.instantdb.com/tutorial. Let us know
what you think!
Author : stopachka
Score : 96 points
Date : 2025-07-16 17:47 UTC (5 hours ago)
(HTM) web link (www.instantdb.com)
(TXT) w3m dump (www.instantdb.com)
| simonw wrote:
| This article is about teaching coding agents to use InstantDB,
| which is "a modern Firebase".
|
| I suggest jumping straight to this document, which is designed to
| tell the agent how to work with Instant but is pretty great
| documentation for humans who want to understand what it can do at
| the same time: https://www.instantdb.com/mcp-tutorial/claude-
| rules.md
| nezaj wrote:
| Thank you for the kind words on the rules/documentation! It was
| definitely an iterative process to figure out how to get good
| results.
|
| We have an llms.txt and llms-full.txt (~9k lines) which
| contains all our documentation. Feeding these to the claude
| didn't get great results, it was just too much information.
|
| We manually compressed our llms-full.txt into a rules file
| (~1.5k lines) which declared the API upfront and provided
| snippets of how to do different things with callouts to common
| examples. This condensed version did better but would cause
| Claude to make subtle mistakes.
|
| Looking at the kind of mistakes Claude made, it seemed like a
| human could make those mistakes too (very useful feedback for
| us to improve our API ). We thought "what's one of the smallest
| fully contained examples we can make that packs a bunch of info
| on how to use Instant?" That would probably be useful for both
| a human and an agent. And indeed it seemed to be the case.
| arscan wrote:
| > Looking at the kind of mistakes Claude made, it seemed like
| a human could make those mistakes too (very useful feedback
| for us to improve our API ).
|
| This is something we've found for our API -- just having LLMs
| attempt to use it helps us identify things that we haven't
| documented well or placed enough emphasis on (for things that
| are critical but are non-obvious or may be drowned out by
| other less important information). Improvements that help the
| LLM tend to be good for developers too.
| stopachka wrote:
| Yes. Fun fact, Instant got the `create` method because of
| how many times LLMs hallucinated it.
| DaiPlusPlus wrote:
| Life imitates art: afar you're describing there is
| basically _The Secret_ (I.e. if I wish hard enough for
| something then eventually it will come true), except it's
| LLMs that get wish-fulfilment, not us.
| CartwheelLinux wrote:
| I'm saving all of these articles for the next time we go through
| the "AI (LLMs) is going to change the world," cycle.
|
| The systems we use can only be as smart and intuitive as the
| people who prompt them.
|
| On top of it, this (LLMs) is not AI, not even close, if anything
| they are glorified prediction systems that require human
| prompting.
| lenerdenator wrote:
| > On top of it, this (LLMs) is not AI, not even close, if
| anything they are glorified prediction systems that require
| human prompting.
|
| _puts in retainer; pushes glasses back up bridge of nose_
|
| Technically schpeaking, what you're talking about is the
| difference between weak AI and strong AI/artificial general
| intelligence (AGI). AGI is the kind of AI that has reached
| human levels of consciousness. We're not there yet. Personally,
| I hope we don't get there, but I'm not the one in charge, so
| _shrug_.
|
| You can do a lot with glorified prediction systems that require
| human prompting. Actually, they are arguably more valuable than
| AGI because you can more easily communicate and utilize their
| value proposition. People don't need a machine that wonders the
| same stuff they do; they need something that does a specific
| task in lieu of their own effort.
| CartwheelLinux wrote:
| Haha. You're 100% correct in the AGI/AI thing. I'm just sick
| and tired of every article being about AI, it's great people
| but we can't stop innovating and attending to other areas of
| technology.
|
| >You can do a lot with glorified prediction systems that
| require human prompting >People don't need a machine that
| wonders the same stuff they do; they need something that does
| a specific task in lieu of their own effort.
|
| This is the problem with our current revision with AI; the
| way I see it those two are in conflict with each other. In
| lieu of their own effort, the way a vast amount of the would
| be users think, is "without promoting" which would lend
| towards AGI than AI.
|
| >Actually, they are arguably more valuable than AGI because
| you can more easily communicate and utilize their value
| proposition.
|
| To you and I this might be true, but to your average non-
| techie I don't think it's quite as true as you would like it
| to be.
|
| Short term it is very true, everyone sees the value until you
| realize it's inherit limitations and the 'shiny, wears off
| 0x457 wrote:
| You're saying it like this is first time AI changes its
| meaning in marketing. People used to market "smart cycle"
| in dishwashers as AI.
| CartwheelLinux wrote:
| In the dishwasher's defense it is pretty smart compared
| to an LLM.
| sidewndr46 wrote:
| Given the level of disruption we'd see if a company reached
| AGI, wouldn't they be incentivized to somehow hide it? They
| could just use said AGI to produce inferior versions of
| itself, each one iteratively a little bit better than before.
| xnx wrote:
| > On top of it, this (LLMs) is not AI, not even close,
|
| Do you think that the LLM/AI tools today are better than those
| from 2 years ago? Do you think the LLM/AI tools in 2 years time
| will be no better than the ones we have today?
| dmitrygr wrote:
| False equivalency. Faster and faster stochastic parrots !=
| intelligence.
| iwontberude wrote:
| You actually answered their question by reducing two years
| of LLM improvements to a factor of speed.
|
| Interpreting your non-response: No, two years have not
| improved things and two more years will not either.
| debarshri wrote:
| If we achieve super intelligence, agents will be shipping
| themselves.
| ASalazarMX wrote:
| Then they will be founding their own startups, and if
| successful, they'll invest in each other's startups.
|
| And every one of them will be ads.
| air7 wrote:
| but who will buy the advertised products? and with what
| money?
| mhog_hn wrote:
| Agents with their agent money - get ready for new legal
| structures and a bifurcation of the economy: agentic and
| human.
|
| Who knows...
| TeMPOraL wrote:
| A separate, self-contained economy.
|
| A Disneyland with no children.
|
| Moloch.
| debarshri wrote:
| May be agents will reproduce small models
| debarshri wrote:
| Agentic democracy
| bluefirebrand wrote:
| This is an implementation detail they'll figure it out as
| they go
| jerf wrote:
| Any finite intelligence will have limits and a "complexity
| budget". I know I see a lot of people assuming AI will just be
| able to do anything, but they can't escape the limits of being
| finite. An AI will benefit from a well-packaged library in a
| similar way to what a human can, though they may have
| meaningfully different preferences on what it should look like.
| physix wrote:
| Read the comments here so far and I find that they are absolutely
| right to offer an AI layer that speeds up building apps on their
| db.
|
| Once built, the solution is plain-old-runnable-code (PORC :-), as
| long as the business logic implemented doesn't exit to LLM. So I
| don't fret so much about the AI hype story here.
|
| For anyone starting off building with new tech, an AI assistant
| is really helpful.
| hoppp wrote:
| Just dont connect an agent to a pay per query database, unless
| you want to risk getting large bills.
|
| Make sure the agent knows how much it costs to query
| nezaj wrote:
| In this case the cost per query is zero!
| sails wrote:
| > Traditionally, end-users were non-technical and would be stuck
| with whatever the application developer gave them. But now every
| user has an LLM too.
|
| Interesting point.
|
| I keep coming back to the idea that users could request changes,
| and they could be experimentally deployed immediately.
| stopachka wrote:
| Thank you. There was a lot to extensions that was bit of scope
| for the essay, which I would love to go deeper on in later
| writing.
|
| Some open questions I had as I thought through extensions:
|
| We talked about the data abstraction side: when you expose
| data, it's easier for end-users to build extensions. But there
| are questions on UIs and data modeling.
|
| UIs: How cool would it be agents could "enter" into
| applications and change the UI? In one sense this hard, but at
| least a demo feels in reach. What if an app exposed the UI
| components that it was built out of? This would let the agent
| remix them.
|
| Data modeling: Exposing data works, but what if users want to
| store extra information? Maybe each user could spin up their
| own separate "extra" database.
| gregsadetsky wrote:
| What other MCP-compatible tools are people using to ship/deploy
| software? Is there anything AWS-compatible that people like/use?
| Something for self-hosters? Anyone letting their agents ssh into
| machines..?
|
| I suppose that most deployment/devops is done using existing git
| push workflows and IaaC. Has anyone had good experience with
| LLM/agent-compatible tools?
| jamest wrote:
| I built an app (HN Clone, of course) with Instant's MCP hooked up
| to Claude Code.
|
| The experience was brilliant.
|
| Pros:
|
| + Fast
|
| + Easy
|
| + "Vibe coding on steroids" basically
|
| + The sense of 'wow' that comes very rarely with new tech
|
| Cons:
|
| - It used Instant as the database/backend, but I wasn't sure what
| it had done / how exactly it worked and had to spend a bunch of
| time asking Claude + reading the code to get it. It seemed
| reasonable, but if I were doing a prod system vs a PoC, this is
| where the time would be spent. ("Vibe coding lets you create tech
| debt 10x faster")
|
| Net-net: This is the way for prototyping / validating. This is
| probably the way for production systems in N months too once the
| toolchain + agents get better.
| croes wrote:
| If they get better. At the moment the progress is on the
| toolchains because the LLMs progress as such slows down because
| of the lack of training data
| achierius wrote:
| Would you mind sharing the code, as well as prompts if you're
| comfortable? I'm trying to sample anecdata to help re-baseline
| my intuition on these things.
| jamest wrote:
| I don't have the prompts, but here you go: -
| https://the-inference.vercel.app -
| https://github.com/jamestamplin/instant-test
| lyime wrote:
| Have you tried Convex?
| ar7hur wrote:
| I've been using InstantDB for two projects for one year and it's
| awesome.
| rylan-talerico wrote:
| Love Instant. Great team and product. Congrats on the launch!
| proxy9 wrote:
| Fun software but the only issue with Instant is their pricing.
| Once they gain adoption, I expect them to significantly raise
| their rates, I can seen them charging over $1 per GB easily. And
| like with any vendor lock-in, you're stuck paying whatever they
| decide to charge. Observe with caution I'd say
| soulofmischief wrote:
| Vendor lock-in from vibe-coded apps is going to be brutal. It's
| an all-out turf war.
|
| But hey, rewriting the plethora of vibe-coded long tail* apps
| might be a major source of employment in the future.
|
| * small but loyal and profitable userbases
| mjsweet wrote:
| The most frustrating problem I have had with Firebase Studio is
| Gemini 2.5 attempting to create firebase rules... it was
| completely unworkable in my experience - just constant
| permissions errors. I pivoted to Claude Code a few weeks ago with
| Prisma ORM and NEON db running on Netlify. It's been pretty good
| so far. I will give InstantDB a go soon I think.
| jen729w wrote:
| Please don't use a vibe-coded app for anything important.
|
| I use Claude. I like Claude. But I've backed away from having
| Claude _actually write my code_ other than in the most limited
| circumstances.
|
| I caught it copying one of my TS Interfaces, for example. And
| modifying, then using, the copy. So my type-checks pass, yay! But
| wait what?
|
| It wrote a test for a tricky bit of code. The test wouldn't pass.
| So it re-wrote it in a way that couldn't possibly fail, mocking
| all elements inside the test itself.
|
| I'm not anti-AI. But I wouldn't trust anything vibe-coded above
| the importance of, say, Wordle.
| stillpointlab wrote:
| This is a massive business opportunity for whoever owns the
| market.
|
| I have a friend who owns a small/medium sized marketing firm.
| They typically manage social media and advertising for local
| businesses (butchers, plumbers, NPOs, etc.). A major cost center
| for them is dev. They can generally handle developing assets
| (images, videos, text copy) and publishing them (Facebook,
| YouTube, Instagram) but if they need any kind of interactivity
| (even basic forms or CRM-like stuff) they used to hire
| programmers.
|
| This friend is now "vibe coding" the simple interactivity that
| previously they had to outsource. In the last few months he has
| pitched, won and crucially _delivered_ simple apps for a few
| clients. We 're not talking complex web apps, it's mostly CRUD
| forms and basic workflows, the kind you see people go on about
| using n8n on Twitter. He's talking to me these days about React,
| Tailwind, DNS and all of that stuff.
|
| His clients don't know, or care, how he delivers. The local
| butcher doesn't know about "best practices" or whatever. He just
| cares that if someone signs up for his newsletter that he gets a
| notification and that person gets his weekly meat deals email.
|
| His firm is picking up more and more complex projects like these
| and saving a huge amount on costs. Turn-key services that enable
| guys like him are going to reap the rewards.
___________________________________________________________________
(page generated 2025-07-16 23:00 UTC)