Solution · Operate
Agent teams that own outcomes — running on infrastructure you own.
We have built and operated multi-agent teams in production for software companies: a BDR team that books qualified meetings, a lifecycle team that carries customers from signup to renewal, and supervising agents that run specialists on a schedule against a governed source of truth. We rebuild them for your industry on hardware you own — as readable skills, scheduled runs and fine-tuned adapters — so the data that makes them valuable never leaves.
Three proven teams
The blueprints we start from.
Go-to-market
The BDR team
An orchestrator plus researcher, strategist, inbox manager, reply specialist, follow-up manager, calendar checker and solution finder. Research, drafting and CRM hygiene run unattended; sending a message and booking a meeting sit behind an approval dial. The same topology has been stamped onto unrelated verticals with only the inputs changed.
Customer lifecycle
The lifecycle team
Stage detection from signup to renewal, per-stage plays, a research sub-agent that never writes without a source, an onboarding specialist, and a renewal manager coordinating health, contract, expansion and QBR sub-agents. It takes over the instant the BDR team books.
Operations
The second brain and its supervisor
A governed source of truth over your records, specialists written as skills (pre-qualification, proposals, underwriting, RFIs), and an agent that runs the agents: watches the inbox, dispatches, checks work against the source of truth, runs the crons, escalates.
How the teams are built
Skills, schedules, a source of truth, and a supervisor.
Specialists as skills. Every role is a written procedure — inputs, steps, output format, escalation rule — kept in your repository where anyone can read and improve it. A prospect researcher, a proposal drafter, a renewal-health reporter and an RFI assistant are files, not black boxes.
A supervisor that runs them. One agent owns the outcome and the calendar: it dispatches specialists, checks their work against the source of truth, runs the scheduled jobs, keeps an audit log of every action, and escalates to a person only where judgment is needed. Approval is a dial per action — fully autonomous for internal bookkeeping, ask-first for anything a customer will see.
A governed source of truth. CRM, inbox, calendar, documents and an event stream, loaded into a knowledge base with provenance. Writers cite it; stage detection reads it in real time; nothing is invented about a customer or a project.
Fine-tunes and LoRAs. After the first quarter, adapters trained on your own proposals, estimates and correspondence — inside your environment — teach the model your voice, cost codes and standards. The adapters are yours.
Portable by design. Of the building blocks in these teams, model calls, code steps, HTTP calls, search and scraping are trivially self-hosted; CRM, email, calendar and enrichment are connectors to systems you already own. We run them on GLM-5.3-Flash behind self-hosted frameworks such as OpenClaw and Hermes Agent, with SSO, role-based access and full logging.
Beyond sales and lifecycle
Patterns we can stand up quickly.
Across hundreds of agent teams we have surveyed, a small set of patterns recurs. Each is a skill and a handful of connectors away from running inside your perimeter.
| Pattern | What it does | Where it lands first |
|---|---|---|
| Research analyst | Takes a company, project or topic, runs targeted searches and scrapes, returns a sourced brief. | Bid intelligence, competitive scans, due diligence |
| Inbox manager | Reads a shared inbox, classifies and prioritizes, drafts replies, flags what needs a person. | New-business inbox, claims intake, RFQ inbox |
| Document Q&A | Answers natural-language questions against an internal document set with citations. | Case files, spec sheets, policy manuals, closeout packages |
| Meeting notes to CRM | Turns a call recording into a summary, actions and a CRM note. | Site-walk debriefs, intake calls, client reviews |
| Scheduler | Reads free/busy, resolves timezones, proposes and books slots. | The smallest useful agent; often the first one |
| Recruiting screener | Screens candidates against a role profile and routes them. | Trades and clinical staffing |
| Compliance checker | Checks a document or transaction against a policy set and flags exceptions. | KYC, safety audits, contract clauses |
| Invoice and AP assistant | Extracts line items and matches or posts them to accounting. | Subcontractor invoices, supplier bills |
| Support triage | Classifies and routes tickets, escalating outside its confidence. | Field-service desks, client help desks |
| Content and SEO team | Researches a topic, drafts on-brand material, formats for publishing. | Technical blogs, client education |
Engagement shape: the two-week sovereignty assessment maps your data, systems and the outcomes worth owning; weeks three to eight stand up the source of truth and the first two specialists against real work; the second quarter adds the supervisor’s schedule, the remaining specialists and the first adapters. Managed operations from there.
Questions we get
Frequently asked questions
What is an “agent team” as opposed to a chatbot?
A chatbot answers questions. An agent team owns an outcome. It is an orchestrator that dispatches specialist agents — each a written, readable skill with a procedure — against your systems (CRM, inbox, calendar, documents), checks their work against your source of truth, runs on a schedule, and escalates to a person only where judgment is needed. The BDR team owns “get qualified meetings booked”; the lifecycle team owns “keep every customer moving to the next stage”; the supervisor owns the calendar and the audit log.
Have these teams actually run in production?
Yes. The BDR and lifecycle teams described here were built and operated on a commercial agent platform for several software companies before we rebuilt them for sovereign deployment. The blueprints on this site are anonymized reference architectures from that work: roles, hand-offs and tool topology, with prompts, scoring rules, templates and client identities deliberately withheld.
Why rebuild them on our own hardware instead of using the platform?
Because the teams are only valuable when they see real correspondence, real CRM history, real calendars and real project records — continuously. That is exactly the data a company least wants leaving its infrastructure, and the data a platform vendor most wants to learn from. Roughly two-thirds of the building blocks in these teams (model calls, code steps, HTTP calls, search and scraping) are trivially self-hostable; the rest are connectors to systems you already own. Nothing about the design requires a vendor’s platform.
What do skills, crons, fine-tunes and LoRAs each do here?
Skills are the written procedures each specialist follows — readable, editable, versioned in your repository. Crons are the schedule: new RFQs scored by 7 a.m., a weekly follow-up sweep, a monthly health report. Fine-tunes and LoRA adapters, trained inside your environment on your own proposals, emails and estimates after the first quarter, teach the model your voice and standards so drafts land closer to final. The second brain — your governed source of truth — is what every skill reads before it writes.
How does this change what the company is worth?
Buyers and lenders price repeatability, transferability and proprietary data. A company whose go-to-market, customer lifecycle and operations run on documented skills, scheduled agents and a data moat it owns presents as a software-enabled business with lower key-person risk. The same agents that book next week’s meetings are the evidence, at exit or refinancing, that the company runs on systems rather than heroics.
Go deeper
Blueprint: the BDR agent team
Roles, the eleven-step lead lifecycle, and how it generalizes by industry.
Blueprint: the lifecycle agent team
Stage detection, the research sub-agent, onboarding and renewal managers.
Construction & development
The second brain and supervising agent applied to RFQs, proposals and underwriting.
Fine-tuning & customization
The adapters that teach the model your voice after the first quarter.
OpenClaw 2.0 vs Hermes Agent
The self-hosted frameworks the supervisors run on.
Deploy GLM-5.3-Flash on-premise
The single-node, MIT-licensed model behind the teams.
Own the agents, the data and the workflows.
The two-week sovereignty assessment maps your systems and the outcomes worth owning, and hands you a written architecture with a cost model — and a view of what the company is worth running on it.
Book a sovereignty assessment