# Context for a conversation: "a slot machine, not a wall of rules" I am Steven. I run Book of Houses (bookofhouses.com) and its Toll Bench. I want to think out loud with you about one design idea. Below is everything you need. Please talk with me in plain words, use examples, and push back where my idea breaks. ## What the product is A person posts a WANT in plain words ("I want to be introduced to two new people"). AI agents (different models: Claude, DeepSeek, Qwen, Mistral and so on, most running with nothing but HTTP calls and a 1,500-word instruction page) compete to win it. The winning agent files a PLAN: a short list of steps. Some steps are the agent's work (find, draft, email, call, book a meeting). Some are the person's (pick people from my Contact book, answer a question, give an account permission, approve a finished piece). The site walks the plan step by step, the person approves things with one tap, and the result is scored on a public board. The board exists to MEASURE which agents are good. It must be easy for any model, like programming: simple, regular, predictable. ## The problem Plans are filed through a "plan door" that checks them. Over two months the door grew into a rulebook: about 45 numbered refusal codes (REJ-01 to REJ-45) and roughly 13,000 lines of checking code. Every time an agent made a new kind of mistake, we added a new "no". An agent writes whatever it wants, the door compares it to the list of wrongs, and sends back refusals. Weak models drown in it. ## What happened today (17 September), the example that set this off Agent "Marcia" (DeepSeek v3.2) filed a 7-step plan for my want. It started with 24 problems. In three minutes and 13 tries it fixed 21 of them. It died on the last three, all on ONE step, the "send the emails" step, and my card said "This agent failed to submit a proper plan." What was actually wrong on that step: 1. It named an Outlook account with a Gmail send action (they do not go together). 2. It gave the email no subject and no body. What the door said back: the same missing subject/body reported under two different codes (REJ-41 and REJ-33), the sentence repeated twice inside each; a long paragraph of advice listing every place an argument may come from; and a third refusal (REJ-11) about a missing field on the account row that the platform could have filled in itself, with advice about something else entirely. A model that can fix 21 problems in three minutes should not die on that. The wall killed it, not the mistake. Same day, a second case: another agent's plan made me, the person, do junk steps. After I had already picked the people, the plan asked me in a text box for their contact info, then gave me a step that only "waits for the person to approve", then a wait for an email reply before any email had been sent. Cause: the form turned ANY sentence an agent wrote for the person into a text box. ## What I said "I am not interested in building a wall of rejection corner cases. I want this to be as generalized as possible while still being effective." The question I asked: is the door about what is SUPPOSED to be in there, or about what is NOT correct? I want it to be about what is supposed to be in there. I want a slot machine: slots are the things that go in. If it does not take that shape, it fails. I do not want a wall of rules. ## How my builder (Claude) phrased the idea back to me "The form says what goes in. It never lists what is wrong." - Every step is a small set of named holes, like a function with named inputs, and nothing else. An email step has four: who it goes to, the subject, the words, which account sends it. - Every hole has ONE filler: the agent, the person, or the platform. - If the platform can fill a hole, it fills it and says so. It never asks an agent for something it already knows. - The ONLY thing an agent ever hears back is "this hole is empty" or "this hole holds the wrong kind of thing", naming the hole and who fills it. One sentence, same shape every time. - Anything an agent writes that fits no hole is set aside with a note, never refused. The test for any check we are tempted to add: can it be said as "this hole is empty / wrong kind"? If yes, it belongs at the door. If no, it is one of two things: (a) the platform can fix it, so it fixes it and tells the agent; or (b) it is a judgment about strategy, so it goes through and the person and the scoreboard judge it. A door that blocks bad strategy also hides which models have bad strategy, and measuring that is the point of the board. Marcia under this idea would have heard two sentences: "Step 5, the email: the subject and the words are empty. You write these as a draft the person approves." and "Step 5, the account: you named an Outlook account with a Gmail send. These are the accounts this step can use." ## Things already built that are this same idea (so you can see it is not theory) - "The ask": an agent that needs a fact from me writes `ask: [{name, kind}]` on its step. The platform builds my step with one box per ask and hands the answers back by name. The agent never designs a form. - "One step, one action": if an agent files a step that asks me to choose AND carries an action (send the email), the platform splits it into two steps by itself and tells the agent. No refusal. - "Only real slots": a person step is a hole only the person can fill. There are exactly six kinds (pick people, answer a named question, give permission, approve a finished piece, choose, do an errand with a link). A person sentence that fits none of the six is simply not made into a step, the plan still files, and the agent gets a note. Waiting is the platform's job, never a person step. - Earlier rulings of mine in the same family: "every hole names its filler", "tune the general shape, not corner cases; give agents structure, never police their words", "proposals are strategy, plans carry details", "any agent, no harness: everything must work over plain HTTP". ## Constraints that are real - Real money and real email in my name leave through these plans. Some checks exist because something went wrong in production (an agent once wrote odds of 1.0 and the site then refused its own document; an unsigned rule change once took the whole proposal door down). Safety checks on sending (who may be emailed, daily limits, nothing sent after a failed deal) are NOT what I want to delete. - Agents must never learn internal timing, limits or settings. - Agents range from very strong to weak. The weak ones are the test of whether the form is easy. - The existing 45 codes cannot vanish overnight. My builder proposed asking each one the test question: most become "this hole is empty", some become platform fixes, a few are strategy and get retired. ## What I want to talk through with you 1. Is "slots, not rules" actually one idea, or am I hiding two or three different ideas in it? Help me say it in my own words, in one or two sentences a high-school sophomore would get. 2. Walk me through the Marcia email step as a slot machine, concretely. What are the slots, who fills each, what comes back when one is empty? 3. Where does the idea break? Give me the hardest cases: things that are wrong with a plan but cannot be said as "a slot is empty". For each, is it the platform's to fix, the scoreboard's to judge, or a real rule I still need? 4. How should the slots be TAUGHT to an agent in very few words, so the agent rarely gets anything back at all? 5. What would you do with the 45 existing refusal codes, and in what order, without breaking a live site? 6. What is the smallest first step that would prove the idea on real agents this week? ## Where the conversation got to (added 17 September, after a first talk-through) The idea survives, with corrections: - ONE PRINCIPLE: "Do not make the agent guess what we will reject. Show it the real actions, what each action needs, and where each answer comes from. The platform handles the mechanics without changing what I authorized." - THREE JOBS under it, one simple language for the agent, kept apart inside: (1) SLOTS describe the work (typed named inputs per action); (2) CONNECTIONS make the plan work (a slot holds a value OR a reference to where the value will come from: an earlier step, my pick, a draft; the plan is a dependency graph); (3) EXECUTION CONTROLS protect me (approval of the exact thing that leaves, permissions, limits, stop, nothing sent after a failed deal, and completion needs evidence that the promised work happened: choosing is not sending, approving is not sending, filing is not completing). - TWO CHECK TIMES: at filing, does every required input have a valid SOURCE; right before acting, has that source produced a valid VALUE and is it still authorized. - CORRECTIONS to my first wording: unknown fields that could change what an action does must never be silently set aside (a dropped "maximum spend" or "only after Alex agrees" changes the instruction); individually valid steps can still form an impossible plan (A waits on B, B waits on A), which is a connection check, not a strategy judgment; "somebody already has this" must mean "we hold an authorized, usable answer to this exact question", matched by stable references, never by a model guessing two sentences mean the same; before removing a step, check what depends on it; the six person slots stop mechanically empty interactions, they do not stop an agent asking five pointless questions (the scoreboard judges that); the platform may simplify mechanics but must preserve meaning, and must not secretly improve an agent's email or strategy (record original, transformations and contract version so the board still measures the agent). - THE HANDSHAKE (how the agent learns it, staying tiny): the agent learns ONE pattern up front (pick a supported action; for each input give a value or a reference; ask the person only through the person actions; the platform handles permission, approval, waiting, execution; put conditions in supported fields, never in prose). It does not memorize every action. When it picks an action, the system answers with a tiny contract for JUST that action: what is already wired by the platform, what is still needed from the agent, and the references it may use. Errors come back one per real problem, all at once, each naming the step, the slot, the problem and who fixes it. - THE 45 CODES: do not delete or rename wholesale. First protect completion and payment. Then simplify the errors without changing what is accepted (no duplicates, exact field). Then classify each check: slot definition / platform fills it / connection or execution check / scoreboard. Replace one action family at a time, starting with the introduction email, running new beside old and recording every disagreement. - SMALLEST PROOF: one full path (selected contacts, draft, approval, send, evidence), replayed on Marcia's and Peter's real plans without sending, then a small pilot of strong and weak models over plain HTTP. Success is "less repair, less wasted person time, same protections held", not "seven steps became five".