A market-data platform that runs itself: custom AI agents watch it, answer its users, build what they ask for and ask a person when it matters
A digital-asset market-data platform with wallets, run by a small team.
Reviewed by Christopher Clayton, CEO
- Scale
- Agents on monitoring, releases, user feedback and communications; 3 independent capture regions watched every minute; 1 human decision rule
- Engagement
- Since the mainnet launch on 4 November 2026, continuously
The challenge
After launch, a platform like this needs the work that never ends: watching the chain and the data feed around the clock, reading what users say, fixing what breaks and building what they ask for. A small team cannot staff that by hand, and a general-purpose chatbot does not know the product, the release rules or what must never be done without approval. The owners wanted a platform that keeps improving without waiting to be prompted, and that stops and asks only when a decision truly belongs to a person.
A product launch is the start of the work, not the end. The week after this platform went live on mainnet it needed monitoring, bug fixes, answers for users and the features they asked for, all at once. Our team's answer was not a bigger team but a set of custom AI agents that know this one product: its code, its runbooks, its release gate and its rules. They are not a chat window waiting for a question. They work on their own schedule, and they come to a person only when a decision is needed.
What they do is concrete. Watchers check the live site, the data service and the media service every minute and cross-check the three independent capture regions every hour. Agents read the results, investigate what looks wrong and open the next task themselves. Requests and bug reports from the community and from support are collected and ranked, and each becomes a scoped task that is built on a branch, passes the automated release gate and review, goes to the warm standby first and can be rolled back in one step. Two requests shipped in the first week this way: a collection activity feed and market capitalization in both currencies.
The rule that makes this safe is the human-in-the-loop rule. Agents do not spend money, touch keys, change live infrastructure or take any irreversible step on their own. When one of those is needed, the agent stops, writes a single numbered question with a recommendation and waits for a yes or a no. The same rule applies to learning: the platform's news lane proposes changes to its own instructions from what readers found useful, and a person approves each one before it takes effect. Communications on Discord, X and email follow written rules too, and anything about money, keys or a sensitive user goes to a person.
The proof is in an ordinary bad day. On 6 November the indexer stopped on a transaction shape it had never seen. An agent traced the loop in the logs, confirmed that the chain and all three archives were intact, wrote a four-line fix from the running commit and handed the build to a person for approval. The site was current again about 65 minutes after it fell behind, and the fix came with a follow-up to make the indexer mark such a block as partial and raise an alert instead of halting. The build of the platform itself is a separate case study; this one is about running it, and about how little of that running now needs a human to start it.
What our team delivered
- Custom AI agents, built by our team around the platform's own code, runbooks and rules, that work without being prompted: they watch the services, read the logs, check the data against independent copies and open the next piece of work themselves
- Round-the-clock watchers outside the agents: a check every minute on the live site, the data service and the media service, an hourly cross-check of the three independent capture points, and weekly restore tests of the encrypted chat backup
- User feedback intake: requests and bug reports from the community and from support are collected, ranked and turned into scoped tasks, and each shipped change is announced back to the people who asked
- A release path that agents cannot skip: every change is built on a branch, goes through the automated release gate and review, ships to the warm standby first and can be rolled back in one step
- Communications on Discord, X and email handled by agents inside written rules: answer questions, post release notes, acknowledge requests; anything about money, keys, legal wording or a sensitive user goes to a person
- A self-review loop: the curated news lane, for example, proposes changes to its own instructions from what readers found useful, but a change takes effect only after a person approves it; there is no unsupervised self-modification
- A human-in-the-loop rule written into every agent's instructions: money, keys, live-infrastructure changes and anything irreversible need an explicit yes; the agent asks one numbered question with a recommendation and waits
Result
The platform has been live on mainnet since 4 November 2026 and has been improved by agents every day since. Two community requests were built and shipped in its first week: a collection activity feed on 4 November and market capitalization in both currencies on 5 November. On 6 November the indexer stopped on a valid transaction shape it had never seen; an agent found the cause in the logs, made a four-line fix from the running commit, a person approved the build, and the site was current again about 65 minutes after it fell behind, with no data lost because the raw blocks were intact in all three archives. People decide; agents do the work in between.
Get this for your company
Related services with fixed prices and a written scope.
AI agent bot, standard
One bot for a task that crosses two systems or handles varied documents, with exception handling and a review queue for your team.
About 3 to 4 weeks
$5,900 per bot
Microsoft Copilot and AI readiness
Before you turn Copilot on, find out what it will be able to see, and fix the permissions first.
Report within 8 business days
$2,400 per tenant
Scanning of software codebase
Static analysis and secret detection across one codebase, reviewed by an engineer so you get findings, not noise.
Report within 3 business days of read-only access
$500 per codebase
Show 3 more
Discovery and requirements workshop
Workshops with your team to map every system, process, user group and integration, and to agree what the new platform must do.
Written requirements and a system inventory in 2 to 3 weeks
$6,500 fixed fee
Cloud security review
A review of one Microsoft 365, Azure, AWS or Google Cloud tenant against recognized benchmarks.
Report within 7 business days of read-only access
$2,800 per tenant
Cloud management
We run your Microsoft 365, Azure, AWS or Google Cloud environment: access, cost, patching, backup and security settings. Priced per month for one tenant.
Onboarding within 15 business days
$1,900 per tenant, per month
After you purchase, we contact you within 1 hour, any time of day or night, to start onboarding.
No payment is taken online. When we call you within the hour we confirm the scope and send an invoice (ACH, card or crypto).
Three approval tiers for AI agents
Every action an agent can take falls into one of three tiers: it runs on its own, a named person approves it first, or a human always does it. The examples are illustrative; each firm draws its own lines during scoping.
A market-data platform with no single point of failure
Three independent capture points record the chain in three regions, a primary and a warm standby serve visitors behind an edge network with no exposed servers, archives go to write-once storage, and three monitoring layers make sure no single failure can silence the alarm.
Twelve months of availability
Measured availability per site over twelve months against a 99.99% target, which allows 52.6 minutes of downtime a year. Six of seven sites meet it; Singapore misses and is shown in red. Illustrative figures.
Facing something similar?
Tell us about your environment. We will tell you plainly what we would do and what it would take.