Templates Overview

LeafMesh ships 20 ready-made teams, with 263 agents (AI agents, human seats and code steps), 291 entry points and 31 scheduled wake-ups: complete, ready-to-run agent teams for real business functions. Each template is a team of agents, not a fixed pipeline: wires are conditional, workflows can start from any of multiple entry points, and the Manager composes the team for each request.

pip install create-leafmesh

create-leafmesh --list                                  # see all templates
create-leafmesh --describe sales-orchestration          # details for one
create-leafmesh my-pipeline --template sales-orchestration
cd my-pipeline

Every template scaffolds a complete project: agency/ with working agent implementations, main.py, Docker files and a .env. There is no configuration file (2.4.498+): create-leafmesh --start loads the team into the agency's store on the first start, only if the store is empty, and you change it in Studio from then on.

The Three Agent Kinds

Kind	What it is	Cost profile
LLM	Reasoning agent — prompt + model + enforced JSON yields, with deterministic post-processing	LLM tokens
Programmatic	Pure Python — ledgers, math, schedulers, scoring. No LLM call	Near zero
Human	A real person as a mesh node (Studio inbox, Slack, webhook) — work is routed to them by `can_call` conditions, with a timeout	Human time

The mix matters: templates put LLM agents where judgment is needed, programmatic agents where determinism is needed (money routing, ledgers, compliance math), and humans where accountability is needed.

One Role, Multiple Responsibilities

Every agent in a template is a role you could hire — an agent is a person, not a function call. Its prompt sections, module functions, and crons are that person's responsibilities: the payment-risk officer runs the duplicate-ledger check and makes the fraud judgment; the recruiting-ops analyst sweeps thin funnels and keeps the time-to-hire ledger. Agents talk peer-to-peer the way colleagues do, handing work across desks rather than up and down a hierarchy. The exceptions are deliberate: controls — deterministic approval routing, QA independence, audit sinks — always stay separate agents, because the person doing the work should never be the person checking it.

Guarantees in Every Template

Immutable audit — every decision terminates in a hash-chained, write-once audit log
Human-in-the-loop default-deny — customer-facing and money-moving actions are always routed to a human agent first
Multi-entry — workflows start from webhooks, schedules, channels, or manual calls
Direct-call endpoints — every non-human agent is independently API-addressable
Peer-to-peer mesh — agents call agents directly (including cycles); the Manager supervises rather than routing every hop

Cycles must be bounded. A can_call back-edge (a retry loop, or a critic looping to a writer) is allowed, but if its condition can stay true forever the chain runs until it hits the mesh.max_chain_events cap (default 50) and the session is stopped with ChainCapExceeded. The templates bound their loops with a per-session exhaustion counter

a terminal route, and so should yours — there is no once_per_session flag. See Coordination → Chain Cap & Bounding Cycles.

Template Catalog

Template	Agents	Core flow	Headline value
[`sales-orchestration`](sales-and-revenue)	12	lead → enrich → score → route → nurture	Zero leads die in the gaps
[`bd-proposals`](sales-and-revenue)	12	scout/ingest → extract → draft → comply	Win more RFPs, find more
[`quote-to-cash`](sales-and-revenue)	14	RFQ → COGS → discount → approve → quote	Margin protection as code
[`marketing-ops`](sales-and-revenue)	15	brief → strategy → create → comply → attribute	Gated pipeline + AI-search visibility
[`customer-success`](customer-operations)	16	sweep → score → risk → playbook → save	Churn caught months early
[`customer-support`](customer-operations)	14	intake → triage → KB → QA gate → reply	Routine 70% self-resolves
[`engineering-sdlc`](engineering-and-infrastructure)	14	spec → build → test ∥ security → review	Seniors only where it counts
[`infra-cost-ops`](engineering-and-infrastructure)	9	anomaly → right-size → compliance veto → tune	Savings without violations
[`finance-ops`](finance-and-procurement)	17	extract → match ∥ recon ∥ fraud → route → post	Audit-proof AP at machine speed
[`procurement-ops`](finance-and-procurement)	11	order → inventory → supplier → PO → receipt	Big-company controls, small team
[`production-planning`](manufacturing-and-supply-chain)	14	demand → capacity → schedule → release	Routine re-plans self-release
[`quality-yield`](manufacturing-and-supply-chain)	10	inspect → classify → root-cause → CAPA	Recall = checklist, not panic
[`predictive-maintenance`](manufacturing-and-supply-chain)	11	sense → predict → schedule → parts → fix	Beat the part-vs-failure race
[`warehouse-supply-chain`](manufacturing-and-supply-chain)	16	allocate → pick → pack → ship	Carrier savings, shrink found
[`services-delivery`](people-and-services)	18	brief → research → draft → QA → review	Margin and capacity per hour
[`recruiting-ops`](people-and-services)	13	source → screen → interview → offer	Faster hires, measurable fairness
[`bench-management`](people-and-services)	11	forecast → match → allocate → reskill	Utilisation points = real money
`_inspire` (default)	8	greet → process → research → advise	Every core pattern, minimal form

235 agents across the catalog — every one a role you could hire.

The Starter Template: `_inspire`

Running create-leafmesh my-project without --template scaffolds _inspire — the teaching template. Its 8 agents demonstrate every pattern the production templates use:

HITL conversation — greeter ↔ client loop
Fan-out / fan-in — processor dispatches to multiple researchers, advisor merges the results
Race pattern — an LLM researcher races a programmatic fallback
Scheduled agents — a daily status cron
Production reference agent — risk_scorer_agent shows the full pattern set: SQLite ledger, record/report modes, threshold escalation, fail-safe returns

Copy from _inspire when building custom meshes.

Next Steps

Quick Start — scaffold and run a mesh in 5 minutes
Cloud Deploy — push any template to AWS, Azure, or GCP with one command
Mesh Architecture — how peer-to-peer agent teams work
Human Agents — configuring the people in the mesh

LeafMesh — YAML-native multi-agent orchestration platform