You were hired to build the agent layer.
Don't spend year one on scaffolding.
The job descriptions say it plainly: Claude and Anthropic APIs, MCP or similar coordination layers, workflow engines, SOX-aligned guardrails, human-in-the-loop. Artifi is that substrate — a governed ledger, the MCP tool surface, the event runtime, and the approval lanes — so you ship agents in weeks and own the logic, not the plumbing.
One headcount. A platform's worth of scope.
Companies are hiring single architects to build finance agent platforms — and asking them for connectors, orchestration, approval routing, evidence capture, and audit logging before the first agent ships. We analyzed ~105 of these postings: the lonely AI architect and the emerging finance-AI stack. The substrate below is the 80% of that job description that shouldn't be rebuilt per company.
Governed double-entry ledger
Multi-entity, multi-currency, dimension-based. 46 transaction types through a unified posting model. Agents never get raw database access — every write goes through the same workflow gateway humans use.
150+ MCP tools
The full finance surface — master data, transactions, banking, reconciliation, reporting, payroll — exposed as tools Claude (or any MCP client) can call, behind permission and entity-scope checks.
Event runtime + approval lanes
Event-driven agent dispatch, risk-scored green/yellow/red approval lanes, human-in-the-loop request objects, exception queues, retries. The reliability layer your JD calls 'validation, retries, guardrails.'
Audit trail as a first-class output
Every agent run is a tracked instance: input, steps, tool calls, cost, errors, and the approval chain on every write. 'Audit-defensible and explainable' is the default, not an export feature.
What you extend yourself, and what ships through core.
Extend through the designed surface and upgrades never break you. We'd rather tell you the boundary now than let you discover it in month three.
Agents — files, not forks
An agent is a directory: config.yaml (triggers, allowed tools, allowed workflows, execution mode) plus prompt.md. Add your own; the runtime syncs them at startup. Country-specific agents auto-scope.
Routing — data, not code
Downstream event rules wire object events to agents with payload mapping and conditions — created through the same submit gateway everything else uses. Rewire workflows without touching core.
Plugins & skills — MIT, forkable
Country rules, statutory filings, and playbooks live in version-controlled plugins (9 first-party today, open on GitHub). Your logic stays yours, upgrade-safe by design.
Core — through us, for everyone
Need a new object type, executor, or tool? It ships in the core through Artifi, and every deployment gets it. You never maintain a private fork of a ledger. That's the deal, stated plainly.
Source access. Escrow. We deploy with you.
Design partners get read access to the platform source under a commercial agreement, a source-code escrow arrangement so vendor risk is answered contractually, and a dedicated deployment — operated by us, on our infrastructure or yours. Your engineers evaluate real code, extend through the designed surface, and shape the core roadmap. Security model: ar-ti-fi.com/security.