Why the Architecture Matters
It is easy to demo a one-message AI outbound flow. "AI sends 100 cold messages overnight." That demo is what gets accounts flagged on LinkedIn, gets prospects mass-unsubscribed, and produces no pipeline. The difference between a working custom AI sales agent and a flagged account is architecture.
A real sales agent has structure: rules it always obeys, a qualification pipeline that kills bad leads before they cost money, a per-tenant memory of what has worked, and a planning phase that runs before any message ships. The architecture is the difference between automation that works and automation that destroys trust.
We are walking through the pattern explicitly because we want buyers to understand what is happening inside a Sentie custom AI agent that handles sales, and because we want potential operators to recognize this pattern when other vendors do not have it.
Layer 1 - The Input: A Single Campaign Prompt
The operator's input is one sentence per major step, written like a brief to a senior SDR:
"Find the people who engaged with this LinkedIn post, qualify them against our ICP, test the new angle in the outreach, send 10 messages."
That is the entire campaign as a prompt. No tool-hopping between scraper, CRM, message composer, and sender. The custom AI agent reads the brief and the next three layers below tell it how to execute.
The input layer matters because it is the operator-facing surface. If the operator has to compose 14 sub-steps to get one campaign out, the agent has not actually compressed the work. If the operator can describe the campaign at the level they would describe it to a teammate, the agent has done its job.
Layer 2 - The Constitution: Rules Every Campaign Obeys
A plain-text markdown file (often called CLAUDE.md or constitution.md or substrate-rules.md) that defines defaults applied to every campaign:
- "Default to plan mode. Never execute without showing me the plan first." - "Never send a message without showing me the draft." - "Exclude existing customers, partners, and competitors from any outbound." - "Cap LinkedIn connection requests at 25 per day per inbox." - "Cap email sends at 100 per day per sender." - "Tone: direct, helpful, no emojis, no exclamation marks."
Write once. Every future campaign inherits the rules. The constitution is the throttle, the brand voice, and the legal compliance layer all in one file. It is also editable by the operator without engineering help, because it is plain text.
This pattern - the markdown constitution file - is showing up everywhere in 2026: [Karpathy's CLAUDE.md repo](https://github.com/forrestchang/andrej-karpathy-skills) hit #1 on GitHub at 44K stars, Anthropic Skills use it, and Sentie's own [Business Brain substrate](/blog/what-is-a-business-brain) uses it per-tenant. The .md file is becoming the programming language of agents.
Layer 3 - The Skills: One File Per Action
Each skill is a container for exactly one action the agent can perform. Scrape LinkedIn post engagers. Qualify a lead against an ICP. Push a record to the CRM. Send an outreach message. One skill, one action, one verifiable output.
A skill bundles three things:
1. The execution rules for that action 2. The connection to the external environment via the relevant API 3. The SDK reference (link to the actual SDK documentation) so the agent does not invent calls
The SDK reference is the non-negotiable part. Without it, the agent writes plausible-looking code that fails silently on the third call. With it, the skill works first try.
Sentie's substrate ships with 248+ skills out of the box. Adding a new one for a specific customer workflow is a Success Manager request, not an engineering ticket.
Layer 4 - The Memory: Everything Specific to You
Per-operator persistent context. The agent reads this on every campaign:
- Your ICP definition (the granular version, with disqualifiers) - Your value proposition (the version you actually say in real calls, not the website version) - Outreach hypotheses you have tested and which ones actually worked - Random ideas you sent yourself at midnight that you do not want to lose - Call transcripts that taught you what prospects really care about
Memory is the layer that compounds. After three campaigns, the agent has accumulated enough specific knowledge about your business that the fourth campaign is materially better than the first - same agent, same skills, richer memory.
This is also what makes switching vendors hard over time. The memory layer for a sales agent is a year of learned patterns about which messages convert in your specific market. That is not portable as a CSV.
The 7-Gate Qualification Pipeline
Inside the agent, every lead runs through seven sequential gates. The structural rule: **cheap gates first, expensive gates last.** Bad leads die before any paid API call runs.
**Gate 1, Dedupe.** Has this lead already been touched? If yes, drop. Double-touching a prospect is the single most damaging mistake in cold outbound - it kills reply rate and trust simultaneously.
**Gate 2, Pre-qualify on headline and title.** Read the lead's LinkedIn headline and current title. About half your scrape volume gets stripped here for free - no paid API call needed.
**Gate 3, Exclusion list match.** Existing clients, partners, team members, competitors. The disqualification layer protects CRM hygiene more than any positive criterion does. It is also the layer most teams forget to build, which is how you accidentally cold-email your own colleagues.
**Gate 4, Profile fetch.** Now the agent calls into the LinkedIn or LinkedIn-equivalent API to pull the full profile. This is the first paid call in the pipeline, which is exactly why it is gate 4 and not gate 1.
**Gate 5, Country and role qualification.** Geographic targeting plus role seniority. ICP fit at the human level. The agent applies the seniority rules from memory and drops anyone outside the band.
**Gate 6, Company selection.** Sector, headcount, industry, "best company from experience" pattern matching against your closed deals. ICP fit at the company level.
**Gate 7, Web fetch enrichment.** Industry context, recent company news, competitor moves, fundraising signals. This is the gate that turns a qualified lead into a personalizable lead. Without it, your outreach reverts to first-name templating.
The funnel shape is the point. A small handful of leads survive all seven gates and get a personalized message. The rest die quietly in the disqualifier, which is the behavior you want. The cost of a bad message in your prospect's inbox is much higher than the cost of dropping a borderline lead.
Plan Mode - The Difference Between a Working Campaign and a Flagged Account
Before any campaign executes, the agent runs in plan mode. It reads the constitution, the memory, every skill spec, and the operator's input, then produces an ordered plan:
"I will scrape engagers using the LinkedIn skill, qualify 47 leads through the 7-gate pipeline, draft one outreach variant in your tone, show you the table of 12 survivors, and on your approval send the messages over the next 6 hours within the 25-per-day connection cap."
You read the plan. You catch anything weird. You approve. The agent executes.
Plan mode is the difference between "AI sent 200 messages overnight and got my LinkedIn account flagged" and "AI ran my campaign on the same throttle my best SDR would and I have a full pipeline by Friday." The constitution carries the throttle so you can carry the strategy.
The Hypothesis Loop - Where the Agent Actually Gets Smart
After the campaign runs, the outcome data goes back into the same agent that launched it. The agent already knows the ICP, the campaign history, the call transcripts, and the value prop. It now has fresh outcome data to learn from.
The operator runs a sparring session with the agent:
"Among the 15 people who replied positively, what do they share that the other 85 do not?"
The agent does not just observe the pattern. It proposes the next version of the qualification criteria, ready to drop into the memory file. Then it stops and asks the operator whether the sample size is high enough to commit. That judgment call stays with the human, but the analysis work is done by the agent.
This is the loop that pulls modern custom AI sales agents ahead of last-generation sales automation. Every campaign teaches the next one. The memory layer compounds. The agent that runs your fourth campaign is materially smarter than the one that ran your first - same code, richer brain.
Multi-Instance Rhythm - The Core+Specialized Pattern in Action
A real sales-agent deployment is usually multiple agents running in parallel, each on its own task:
- One agent runs the qualification campaign in progress - A second agent drafts follow-up sequences for leads who replied last week - A third agent rebuilds the scoring rubric based on yesterday's call transcripts
None of them blocks the others. Each pings the operator when it needs context. **This is the core+specialized agents model in practice** - a coordinating agent and specialized agents for specific functions, all running on the same substrate, all reading the same memory.
Sentie's customers experience this as "my Sentie agents handle sales" not "I configured one agent to do sales." The pluralization is correct. Real sales coverage needs a small group of agents working together, just as a real sales team is a group of people with different specializations.
How to Tell If a Vendor's Sales Agent Is Real
Five questions to ask any vendor pitching you a custom AI sales agent. The answers separate real architecture from a foundation model with a frontend.
**1. Show me the constitution file.** If they cannot or will not, the agent does not have configurable rules. Buyer beware.
**2. Walk me through your qualification gates.** Real agents have 5-8 sequential gates with cheap-first ordering. "It just figures out who is a good fit" is not an answer.
**3. How does the agent persist memory across campaigns?** If the answer is "it does not, each run is fresh," the agent cannot learn. You are paying for memory-free automation.
**4. Can I see plan mode before execution?** If the agent goes straight from prompt to action, you have no defense against the bad batch.
**5. What is the throttle?** Real sales agents enforce per-channel rate limits. If the vendor cannot tell you what theirs is, your account is at risk.
Sentie answers yes-with-detail to all five. If you are evaluating us, ask. If you are evaluating a competitor, ask them too.
The Practical Next Step
If you want to see this architecture working on your specific outbound, the free assessment is the fastest path. Your Success Manager configures the constitution to your throttles, sets up the qualification gates against your ICP, wires the skills into your CRM and communication tools, and walks you through plan mode before anything ships.
[Start the free assessment](/onboarding). No credit card, no commitment, real Success Manager runs the discovery.