two contracts for the second step how an agent finds what to follow, and why nothing it finds is trusted. first: reaching into a response. a path can walk fields with dots and step into arrays with [], so: pairs[].baseToken.address can pull token addresses out of every entry in a screener response. deliberately small. enough to reach a field inside a list, not enough to become a query language. anything that isn’t a string or number gets skipped rather than coerced. if an object appears where an address was expected, that’s treated as a mistake, not a value. github.com/gardnsolana/devel… then the part that matters: everything an agent follows came from content served by someone else, so none of it is trusted automatically. every extracted value is validated for the kind it claims to be before anything fetches it. and if it looks like a URL, it goes through the exact same policy as a target URL. the test that matters here: a feed returns: → 169.254.169.254 → a loopback address → a URL carrying credentials none of them get followed. a response can’t smuggle an internal address into something the agent then reads on its behalf. the fan-out is capped too, regardless of what a plan asks for, so a run stays predictable in time and cost. every attempted target appears in the result as ok or failed. nothing disappears quietly. if a follow fails, it’s recorded. if a whole step returns nothing, that step is treated as broken rather than being mistaken for a valid “nothing found” result. github.com/gardnsolana/devel… also fixed today: a plan carrying a follow-up is now always treated as an endpoint agent, because only an endpoint read gives you a payload you can path into. getting that wrong was what made an agent meant to screen a feed ask for one token address instead. found it, fixed it, added the contract. 22 contracts. 167 tests green. 🌱

day four wrapped 🌱 a full build day on GARDN Intelligence, ahead of tomorrow's release. what landed: → agents can now read the open internet with one public https endpoint, reasoned over → reasoning tiers: fast, balanced, deep different models, different depth, real measured cost → reasoning now runs even when nothing matched, so a no-match explains itself → every run reports what it observed, not just what a rule asked about → severity on every verdict: informational, worth watching, act now → the Conservatory: design an agent by talking to GARDN, and prove it without leaving → the runtime clock is now ground truth for the reasoning engine → page shells and bot challenges are named as such instead of burning a run → 9 commits on the app, 5 contracts made public the important part is what we found on the way. GARDN Intelligence did not work. not "was rough" did not work. the tool schemas sent a field the model API rejects outright, so every planner and reasoning call failed before it ever reached a model. that code predated today. the tests mocked the API, so nothing caught it. we only found it because we stopped trusting the test suite and ran the real thing against a real key. it would otherwise have shipped tomorrow, live, in front of everyone. the same afternoon turned up two more: agents were reading only the first 12k characters of any response, which silently reduced a thirty-item feed to one usable entry. and the reasoning engine, never told the date, was reading today as the future and calling healthy runs stale. all three are fixed. all three are the kind of thing you only find by using your own product properly. five contracts are public tonight, including the one that matters most: hard rules gate the verdict, and the model can withhold a match but can never create one. a reading agent that can be argued into firing is worse than no agent. tomorrow: release 03. GARDN Intelligence. day four done. 🌱

first look at GARDN Intelligence 🌱 not every task needs the same amount of thinking. a quick lookup shouldn’t run like a deep investigation, and a more complex job shouldn’t be forced through the same lightweight path. GARDN Intelligence starts changing that. you’ll be able to describe the agent you want in plain english, shape what it should do, and choose how deeply it should reason before it acts. fast when the job is simple. deeper when the job demands it. the important part is that the reasoning still sits on top of the same verified data, testing and evidence trail already built into GARDN. tomorrow, users get direct control over that choice. release 03 is getting close. more work ahead.