6:42 a.m. The port is closed. Most companies start a group chat and a spreadsheet. We built the other version: your operation as code, agents acting inside limits, one decision waiting on a named person to sign. The port is still closed. The operation isn't. Contact us to see it live!
Our favorite AI conversations are with skeptical IT leaders.
Bring us your biggest operational headache.
4 answers. 3 minutes. A real solution.
No endless slide decks. No AI buzzword bingo.
Just a chance to build something together that actually works.
We’d rather show you than tell you.
An agent working from rules nobody wrote down is a very confident guess.
Two minutes on how we write the rules down first, then put digital workers inside them.
@serlekan The line where the brakes engage should not be in the agent at all. It is a write permission the agent credential does not hold, so the proposal must pass a second identity to land. If the brake lives in the prompt or the loop, there is no brake, just a request to please stop.
@weijinnnnnn The variable is what the exception arrives with. "Invoice failed match" is a hunt. "PO 100, received 96, invoiced 100, tolerance 2%" is a decision. Same exception, same expert. Review time per set of books drops when the agent packs the evidence, not when it touches more.
@sharoonthomas Test-before-live is the bigger change than the plain English. Rule builders never showed the diff. One thing English has to settle: when two instructions collide (a 20 kg item in an order that should stay zone-grouped), which wins, and does the test flag it or quietly pick?
@MichaelGannotti The never-runs direction is the one that hides. A HOLD nobody clears looks identical, a week later, to an agent that did nothing, except the item it was holding is now late. Does HOLD carry a clock, and when it runs out does the card escalate or just stay stamped?
@v_shakthi Execute is really two classes. Cancelling an order still in the pick queue and cancelling one already on the truck are both Execute, and only one undoes quietly. The split that matters is whether a counterparty has seen it. When the timeout expires, does it escalate or just sit?
@A_Intimidating Further than the tooling suggests. Per-agent credentials and signed action logs exist today. What is missing is habit: most shops still provision an agent like a new hire on a shared login. It turns standard the first time an auditor asks which agent did this and nobody can say.
@A_Intimidating Two things, both boring. Before an agent touches anything, write down what it may do alone and what it must hand back, one sentence each. Then run one workflow to a measured result before starting a second. Most cancellations are pilots that skipped both and scaled ambiguity.
A system of record shipped 260+ agent tools and 490+ APIs this week. An outside agent can now read, write and go live without anyone opening the product.
Access was never the hard part. Knowing which of eighty retailer rulebooks governs this pallet is.
An agent without a harness doesn't fail cleanly. It fails a different way each time. One measured workflow went from ~30% reliable to ~97% once the harness got built in.
https://t.co/xcfhpIN1mz
@stretchcloud Logs miss one more thing besides causation: the rule version. An agent posts an invoice inside a 2% tolerance, the tolerance drops to 1% next week, and the trace shows a correct action against a rule that no longer exists. Store the policy in force next to the action.
@MichaelGannotti Sovereignty and cost point the same way once the layer around the model does the enforcing. In our published bench an open-weight model inside the same run-time controls matched the frontier model end to end at roughly a tenth the cost per token. What stays local first?
@Pksingh900 The one that breaks first is knowing which version is official. Two agents update available-to-promise on one SKU, the shared layer holds both, the order agent reads the older one and promises stock that already shipped. Single writer for state, timestamp check on every read.
@puneetmehtanyc The interceptor in front of the ERP is the right call, with one AP-side catch: many ERPs accept the write and post it in a nightly batch, and some bounce overnight on a closed period or a blocked vendor. Agent done and ledger done are hours apart. Do you reconcile to the batch?
Every AP team runs a three-way match tolerance. Ask who set it and when, and the room goes quiet. Hand that rule to an agent and a decision nobody remembers making now runs on every invoice, all month. Automating the rule is easy. Checking it is still right is the job.
@gregisenberg The freight claim example is dead on: the harness also has to know when it's out of pattern and stop, not guess. A handful of denial reasons cover most of a carrier's rejects. The exception queue is where the value and the risk both live.
@Goura_vk None of those alone. What earns it is a graduated system: full oversight until the pattern holds, then autonomy for routine work while edge cases still escalate. Trust isn't one threshold, it's a level you raise workflow by workflow as the log backs it up.
@thoughtson_tech@Zephyr_hg The freight version of this is carrier rate tables, same trick: patient parsing until the overcharge falls out. We caught $50K that way once. The audit problem is identical too: a closed model's finding can't be verified without paying that vendor to check its own work.