Legacy Builder #lockedin

progress update 🌱 scheduled runs. an agent that runs itself. up to now every agent has been something you press. build it, prove it, hit run. fine for testing, useless for the thing people actually want which is to not be watching. this one wakes up on its own. what's on screen that's one agent, on a five minute cadence, with 8 runs behind it. i didn't trigger any of the recent ones. i set the cadence and left it alone. NEXT RUN counts down. when it hits zero the agent goes, and the number goes up by one. what a run actually is each one is the full loop, not a ping: ✓ fetches the source live - real dexscreener data, not a cached copy ✓ reads what came back and reasons over it ✓ decides whether anything crossed the line you set ✓ writes down what it saw, with the numbers, as evidence you can open later ✓ books its own next run before it finishes so 8 runs isn't 8 checks. it's 8 recorded observations of the market, each one timestamped, each one openable, each one showing what the data looked like at that moment and what the agent concluded from it. that's the bit that matters. an agent that runs unattended and keeps nothing is just a cron job. the evidence is what makes it worth anything later. the hard part wasn't the timer the timer is easy. everything around it isn't. what happens when the source is down at 4am. what happens when the same tick fires twice and tries to run the agent twice. what happens when a worker dies halfway through. what happens when an agent has been quietly failing for six hours and nobody's looked. that last one is the real risk. an unattended agent is the only thing in the system where a bug doesn't show up as one bad result you're staring at, it shows up as everything silently broken overnight. so an agent that keeps failing now takes itself out. five failures in a row and it pauses, records why, and tells you. better a stopped agent you can see than a broken one you can't. why this unlocks the next part you can't sell someone an agent they have to keep pressing. everything we've shipped - reasoning, keys, agents that find their own source - has been building something worth leaving alone. this is the piece where you actually get to leave it alone, and it's the last one before an agent is worth owning rather than just running. next: publish it, let someone else fork it. not on the live site yet. finishing checks first. live this evening 🌱

progress: agents that take steps 🌱 got the first version of this working today. one agent read a live pair feed, pulled every token address from the response, then checked each token on-chain for: → mint authority → freeze authority → holder concentration then compared the results across the whole set. two data sources in one run: an off-chain feed to find candidates, then Solana itself to verify them. not two agents. not two clicks. you configure none of that manually either. you can say: “screen solana pairs, then check each token’s authorities and tell me which are risky” and the agent works out the first read, where the addresses sit in the response, what needs following and how many items to check. the fan-out is capped, every followed value is validated before it’s read, and failed follows show as failed rather than disappearing. until now, a GARDN agent read one thing and answered. this is the first version that goes and checks what it found. still building on it. next: sources that need a key.

two contracts are public tonight. agents can now read the open internet, and reasoning has a policy you can audit. endpoint targets. an agent can read one public https url and reason over the response. the url is user-supplied, so it is validated before any request leaves the runtime: https only, no embedded credentials, no loopback, private, link-local or carrier-nat addresses, and cloud metadata endpoints refused outright. redirects are followed manually and every hop is revalidated with the same function, so a public url cannot bounce you somewhere internal. reads are GET-only, capped at 64KB, and abandoned after 10 seconds. reasoning policy. the part worth reading twice. analysis and verdict are deliberately separate. analysis is always produced, including when nothing matched, so a no-match carries an explanation instead of a bare failed rule. the verdict stays rules-first: hard rules are evaluated outside the model, and the model can withhold a match the rules allowed but can never create one the rules refused. that asymmetry is the safety property. observations contain attacker-controllable text on-chain metadata, fetched web pages and a reading agent that could be talked into firing is worse than no agent at all. so it can't be. both contracts ship with the tests that prove it. github.com/gardnsolana/devel… github.com/gardnsolana/devel…