Enterprise AI transformation at startup speed. 🟪 https://nitter.cf/t.co/O9yVkOtwhM
Subscribe to our newsletter →
Joined May 2026
- Tweets1.6K
- Following556
- Followers13.3K
- Likes3.4K
Pinned Tweet
Some patterns we're seeing across our engineering clients that tell you where AI adoption is actually heading:
1. Companies are done spending blind on AI. A few months ago, most were approving unlimited API budgets for OpenAI and Anthropic with no tracking on where the tokens went. Today they're asking us to build governance layers and monitor every dollar. One client had three teams running up five-figure token bills before a single production agent shipped. Now they want dashboards that show exactly what's AI-generated, what it costs, and whether it's actually faster than adding headcount. That means more diagnostic work for us and meaningful savings for them.
2. Companies are accepting the real timeline. A few months ago, leadership expected to go AI-native over a few workshops and a handful of tool subscriptions. Now they're calling us because the codebase is seven years old and can't absorb automation without a full foundation phase first. Their developers are using AI at 60-70% adoption, but nobody's tracking what's generated vs. human-written. QA hasn't been touched. There's no governance. Going AI-native means rethinking org charts, handoffs, and process ownership across every department. Not a software purchase.
3. Companies are walking away from the big firms. The Accentures and McKinseys are all pitching AI transformations. Our clients keep showing up saying the same thing: "We paid seven figures for a strategy deck and nothing shipped." This is why firms like ours have a market. We operate at the engineering layer while having the business sense to map a process end to end, then build the agents to automate it. A year ago we had to explain this on every first call. Now prospects open the conversation with it.
In the last 90 days we've taken calls with PE firms, law firms, healthcare companies, VPN providers, construction platforms, and retail tech businesses. Some are direct competitors with each other. Nearly all came through referral.
AI SaaS won't transform your company. Giving every developer a Copilot subscription and calling it a strategy won't either. The only path to real operational ROI is combining a team that learns your business processes on the ground with engineering capacity to automate every manual workflow worth automating.
We built Limestone Digital around this thesis ten years ago. AI made it urgent. If you're looking to transform engineering delivery or automate business processes with production-grade agents, visit limestonedigital.com.
Limestone Digital retweeted
We built a fleet of AI agents that talk to patients about bills they owe.
The most important agent in that system never talks to anyone. It only reads.
A US healthtech startup gets paid only when a patient's balance reaches zero and the payment reconciles against the right charge. Its agents call, text, email and mail patients on the provider's behalf. Get it wrong and the provider loses a patient.
The team spotted the risk early. Any single email from the agent could look fine. Read a whole thread of 10 in order, and you could see where the model had lost context.
Most teams test an agent one answer at a time. Patients get the whole conversation.
So we added a layer on top. An agent reads every thread, every message in order across every channel, the way a quality manager would. Did it answer the question? Did it say something it shouldn't have? Should this have gone to a person?
By late September it had reviewed 5,135 threads and raised 835 findings.
Then we closed the loop. Separate judges go through the transcripts and propose fixes, 261 so far. A person triages every one. For every fix marked done, another model goes back to calls made after the change and rules on whether the problem actually went away.
Stats on real volume: 61,246 calls, around 97% of patient conversations finished without a person touching them, and 18,194 balances driven to zero.
Of 949,563 messages the system planned to send, it pulled back 658,657 before they went out, mostly because the patient had paid, disputed or changed status.
If you're looking to put AI agents to work in your company, book a call with us.
Velocity Core by Limestone Digital turns engineering tickets into reviewed pull requests, on your own infrastructure.
We built it to move work through your existing software delivery process, with clear rules for what the agent can do and when a person needs to step in.
Add a label to a ticket, and the workflow starts:
→ Check the task before writing code. Admission checks catch missing requirements. If the ticket is unclear, the run stops and asks for an answer.
→ Build with the right context. The agent uses your repository, engineering rules and task context to write a specification and produce a patch inside an isolated sandbox.
→ Validate the change. Lint, type checks and fast tests run before delivery. Failed checks enter a repair loop with limits on steps and spending.
→ Inspect the diff before pushing. The diff guard checks for deleted tests, protected-file changes and secrets. Critical changes wait for human approval, then go through validation again.
→ Work through CI and review. Velocity Core opens the pull request, handles failed builds within configured limits and responds to review comments. A person always makes the merge decision.
→ Keep costs visible and capped. Every run has a hard spending limit. Retries stay inside that budget, and costs appear in the ticket, pull request and dashboard.
The agent holds no credentials. A separate worker handles pushes, and model calls go through your gateway to the provider you choose. There is no Limestone control plane or standing Limestone access.
Autonomy is earned per repository. A repo can start with plans or draft PRs and progress as its results justify it. Your team approves every promotion.
We start with a pilot on live tickets you choose. A Limestone engineer helps install the system, reviews the work and trains your team.
Then we measure what matters: first-pass CI rate, cost per merged PR and reviewer time.
Want to see how Velocity Core would work with your backlog?
Book a discovery call below.
Book a discovery call with us: limestonedigital.com/book-a-…
Limestone Digital retweeted
follow this account if you want to learn enterprise AI without the hype:
"We basically ended up vibe coding a bunch of slop."
This is what a co-founder of a company that builds software for government agencies told our team.
Under their earlier technical leadership, the team built with a spec-driven AI tool. On paper they did everything right. Specs first, agents building from the specs, tests generated automatically.
Their two senior engineers walked us through what they found when they looked closer.
The specs were written by AI too. "The specs themselves were vibe coded and not properly curated by a human." So the system design decisions inside them went straight into the code. The tool also left piles of generated documents in the repo, and the engineers now plan to delete most of them.
The agent said it was done when it wasn't. "The agent says, yeah, I did the thing, and it didn't do the thing."
The tests passed every time. They were "testing the same thing that's in the code," so they could never fail.
Every check in that process was AI checking AI.
When our team showed them delivery dashboards with commits shipped and tickets closed, their engineers pushed back. "It's not a quality measurement." The co-founder wanted to know how we avoid hallucination, context rot and poor abstraction. Fair question.
The reality is the spec is the one part you can't hand to the model. Here's how our engineers answered it:
> audit the codebase before anyone builds, so you know what the slop already broke
> a person writes and signs off every spec, and the agent only helps draft it
> the spec gets a changelog in the pipeline, so it still matches the feature after the tenth change
> tests come from the spec, never from the code they're testing
> each module gets a short context file, so the agent stops guessing what it is
Then the pilot they asked for. One feature from their customer backlog, not trivial, judged on speed, quality and cost before anything bigger.
AI is an amplifier.
Give it specs nobody read and it amplifies the bullshit.
BREAKING: Claude Code can now run 2 to 3 workers in parallel, have Codex review their code, and refuse to call anything done until it's checked the diff itself.
Subagents were 34% of one heavy user's weighted usage over 7 days in our logs, so it stays small.
Here's the prompt:
Credit first: Vox (Voxyz_ai on X) posted the Herdr version of this pattern on 31 Oct, including the how often x how bad score.
Optional: Herdr, if you want one window showing every worker. The prompt above doesn't need it.
It's open source (Apache 2.0), with Can Celik as top contributor. A local server keeps agents alive, a sidebar shows who's working, done or blocked, and its CLI, hc, starts and steers them. Apple silicon Macs, Linux, Windows.
On Mac and Linux, install adds the app, herdr-core and hc in your home folder, with no sudo and no shell config edits. The app updates itself and sends a daily anonymous usage count unless you turn it off. Its optional skill tells agents to prefer hc over their own subagents.
Treat it as a supply-chain step:
→ Read the install script from herdr.dev/download before you pipe it to a shell. Its checksum comes from the same release server, so it catches corruption, not a compromised server.
→ The desktop app's repo dates from 16 Oct 2026.
→ Get approval before a new agent tool touches client code.
Using it? Have Claude read hc --skill first and run Codex with --no-daemon, as Herdr's manual requires. We haven't run this variant.
Want to put AI to work in your operations next?
Get our free guide: how to pick your first workflow, measure the results and scale what works, with 3 case studies from healthcare, insurance and commercial lending.
thefoundation.limestonedigit…
Limestone Digital retweeted
"We basically ended up vibe coding a bunch of slop."
This is what a co-founder of a company that builds software for government agencies told our team.
Under their earlier technical leadership, the team built with a spec-driven AI tool. On paper they did everything right. Specs first, agents building from the specs, tests generated automatically.
Their two senior engineers walked us through what they found when they looked closer.
The specs were written by AI too. "The specs themselves were vibe coded and not properly curated by a human." So the system design decisions inside them went straight into the code. The tool also left piles of generated documents in the repo, and the engineers now plan to delete most of them.
The agent said it was done when it wasn't. "The agent says, yeah, I did the thing, and it didn't do the thing."
The tests passed every time. They were "testing the same thing that's in the code," so they could never fail.
Every check in that process was AI checking AI.
When our team showed them delivery dashboards with commits shipped and tickets closed, their engineers pushed back. "It's not a quality measurement." The co-founder wanted to know how we avoid hallucination, context rot and poor abstraction. Fair question.
The reality is the spec is the one part you can't hand to the model. Here's how our engineers answered it:
> audit the codebase before anyone builds, so you know what the slop already broke
> a person writes and signs off every spec, and the agent only helps draft it
> the spec gets a changelog in the pipeline, so it still matches the feature after the tenth change
> tests come from the spec, never from the code they're testing
> each module gets a short context file, so the agent stops guessing what it is
Then the pilot they asked for. One feature from their customer backlog, not trivial, judged on speed, quality and cost before anything bigger.
AI is an amplifier.
Give it specs nobody read and it amplifies the bullshit.
Limestone Digital retweeted
BREAKING: Claude Code can now build the eval suite that shows what your next prompt or model change breaks, before a customer does.
When Anthropic tested one change on a sample agent, the break showed up in only 1 of 2 runs. So this runs every case 3 times.
Here's the prompt:
Limestone Digital retweeted
Two owners asked me last month how to get their people using AI. I told them what they'd hear first.
"We're in this constant whirlwind. We're sucked in. We don't have time."
It's unsexy, but it's reality. Your people have the same capacity they had before AI, and now they have a new way of working to learn on top of the job. For a while it's harder than working without AI, so they're not making excuses.
You can't wait for a quiet quarter, because it won't come. At our own company we do four things:
1. Have them ask the agent first: "Here's what I'm doing day to day. What could I do using you?" Curiosity does more here than any training deck.
2. Give them a date and a real deliverable. "By the end of next month, you deliver this report using the coding agent." Use a report they already produce.
3. Have them come back and show you how they did it. You can't pick the use cases for them, because you don't know their day. They do.
4. Take something off their plate while they learn. If nothing comes off, you've told them this is optional.
Then keep asking, every day, with a little bit of patience. I haven't found another way.
Limestone Digital retweeted
50 AI engineering concepts every CTO should be able to explain before an agent touches production:
1/ Baseline metric
2/ Cost of a wrong answer
3/ Reversible vs irreversible actions
4/ Build vs buy
5/ Data readiness
6/ Pilot exit criteria
7/ Context engineering
8/ Prompt caching
9/ Compaction
10/ Prompt versioning
11/ Tool descriptions
12/ Schema validation
13/ Agent harness
14/ Coordinator and workers
15/ Subagent isolation
16/ Idempotent tools
17/ Durable execution
18/ Checkpoints and resume
19/ Stop conditions
20/ Escalation paths
21/ Precision and recall
22/ Regression evals
23/ Offline vs online evals
24/ Data drift
25/ Trace replay
26/ Shadow mode
27/ Canary releases
28/ Model version pinning
29/ Per-run spend caps
30/ Rate limits
31/ Batch processing
32/ Model deprecation
33/ Least-privilege tool access
34/ Untrusted content boundaries
35/ Agent identity
36/ Secrets management
37/ Network egress controls
38/ Third-party tool vetting
39/ Audit trails
40/ Zero data retention
41/ Data residency
42/ Business associate agreements
43/ Model risk management
44/ Decision explainability
45/ Approval gates
46/ Automation bias
47/ Adoption rate
48/ Kill switch
49/ Rollback plan
50/ Who gets paged at 2am
If you can explain all 50 in plain English, you'll spot a bluffing AI vendor in one meeting.
Bookmark this. Find the 5 your team can't explain. Start there before your next release.
BREAKING: Claude Code can now rewrite a legacy module and prove, with a script, that it behaves like the old code.
On Anthropic's COBOL-to-Java test, 5 independent checks caught things the tests missed. The last found 7 more differences and ruled NOT PROVEN.
Here's the prompt:
Before you paste it:
→ open an empty folder for the work, start Claude Code there and install the plugin
→ have Python 3.8 or newer as python3 and the target stack's toolchain. The build and the proof need both
→ bring an engineer who knows how the system builds and runs. Preflight asks 5 questions only a person can answer
→ expect real usage. In Anthropic's runs, rule extraction started 50 to 200 agents, so the prompt scopes it to one module
→ if the old code only runs on a mainframe, collect recorded outputs from a test region or de-identified runs first. Without them, the proof has nothing to compare against
Pilot a small policy, claims or billing module first. Save the one your team avoids until the proof has earned trust.
Want to put AI to work in your operations next?
Get our free guide: how to pick your first workflow, measure the results and scale what works, with 3 case studies from healthcare, insurance and commercial lending.
thefoundation.limestonedigit…
Limestone Digital retweeted
A head of innovation at a PE-backed fund services firm told me what they found in their first month on the job.
Contractors had come in and built AI tools fast. "Vibe coding," in their words. A lot of money went out. What came back was some basic HR automation and, as they put it, "nothing that's production ready."
Their verdict on the contractor spend: "It has not gone well."
We've seen this on a lot of calls this fall. A demo takes a few days now, so a vendor can show progress for months without anything going live.
If you've inherited a pile of AI work like this, four questions will sort it:
1. Where does it run? If the answer is somebody's laptop or personal account, you have a prototype.
2. Who gets the call when it breaks during month-end? If you can't name the person, nobody owns it.
3. What do you test it against? Without a set of correct answers your own team signed off on, you can't tell whether it still works after the next model update.
4. Which number moved? Cost per case, days to close, cash collected. If nobody can name one, you paid for a demo.
This person has more ideas than they can count. They asked me for a few things to get shipped, and in my view that's the right ask.
Taking what somebody already built and making it a reputable, governed, stable thing is a big part of our work.
Limestone Digital retweeted
The CEO of a PE-held company told our team the whole company has a Claude account. "That's not the problem," they said. "It's like giving kids keys to the car."
They added straight away that they don't see their people as kids. Their point was what happens next: "I'm using Claude like Google. I'm answering questions or I'm having it write the emails."
We heard versions of this on three calls in the same month.
A director at a financial services firm told us about a third of the firm has Claude. Their team built plugins for core workflows, and only "a very small handful" of people have used them. The plugin needs a heavy extra step first, and as they put it, "you want to get a capital call out today for your client and it just adds like these hurdles."
An executive at a PE-backed construction roll-up called their AI experience growing but very limited: a few heavy users, and plenty of people who've barely touched it.
All three companies have the licenses. What they have today is individual productivity, which is fine, and it won't show up in the P&L.
I'd do two things this week:
1. Ask each department head to show you one step in a weekly workflow that AI removed. A faster email doesn't count.
2. Test any new AI step on the worst day of the month. If it adds work before it saves work, people on a deadline will skip it, and they'll be right to.
With that said, if you build agents on top of people who don't use agents, you won't get anywhere.