System Manual
Overview & Philosophy
What eClips Workforce is, the autonomic model behind it, and the three ideas that make a fleet of agents safe to run unattended.
What eClips Workforce is
eClips Workforce is an autonomic operations layer for a business: a fleet of specialized AI agents that read the work arriving every day — invoices, contracts, leads, tickets, job applications — route each item by your rules, act through your existing tools, and escalate only the exceptions a human genuinely needs to see.
The canonical unit of work is a run: one document or task handed to one agent, executed as a state machine, and resolved into either an action taken or an exception queued for a person. Everything in this manual — the architecture, the fleet, the reliability harness, the trust model, and the credit engine — exists to make those runs both autonomous and accountable.
The autonomic model
Most automation tools wait for a human to click a button. eClips Workforce inverts that. Agents work continuously against the work that arrives, and humans govern the system the way they govern a team — by setting rules and reviewing exceptions, not by approving every keystroke. We call this the autonomic model: the routine runs itself; human attention is spent only where judgment is genuinely required.
Human governs, machine acts
The division of labor is deliberate. The machine acts: it reads, classifies, decides, and executes through connected systems. The human governs: they author the rules each agent follows, set the budgets, decide how much autonomy an agent has earned, and review the exceptions that land in the triage inbox. A person is never in the loop on routine work, but always in control of the policy that shapes it.
The three ideas that make it safe
Three mechanisms keep an unattended fleet trustworthy. Each is detailed in its own chapter of this manual.
- •
Every run is scored
Each run produces a confidence score. Anything below the auto-approve bar is sent to a human triage queue instead of acting blindly. - •
Autonomy is earned
Each agent carries a trust level per organization. It starts supervised and is promoted only after a track record of correct decisions — and demoted on errors. - •
Everything is metered and audited
Every action is metered in credits, gated by trust, and written to an audit trail, so the system is always accountable and never silently expensive.
The three confidence bands
The product surfaces three bands derived from the agent's confidence score. These thresholds are defined centrally in CONFIDENCE_THRESHOLDS and applied at every routing decision in the orchestrator.
| Auto-approve — confidence ≥ 0.85 | The run proceeds automatically and acts through your connected tools. |
| Warn — 0.65 to 0.85 | The run proceeds but a warning is logged on the trace for later review. |
| Dead zone — below 0.50 | The run always goes to triage and never acts on its own. |
Example · A run in each band
A clean PDF invoice from a known vendor extracts at 0.93 — above the bar, so the invoice processor acts and posts it. A smudged scan extracts at 0.71 — in the warn band, it proceeds but flags the trace. A handwritten note with an ambiguous total scores 0.40 — in the dead zone, so it is queued for a human with the proposed reading attached.
Who this manual is for
This manual is the reference for everyone who operates, integrates with, or evaluates eClips Workforce:
- Operators who configure agents, write rules, and work the triage inbox day to day.
- Administrators who manage plans, seats, per-member budgets, and trust promotions.
- Developers integrating through the public API or building on the MCP and Reach surfaces.
- Security and finance reviewers assessing isolation, auditability, and cost control.
◳ Screenshot
A high-level overview graphic: work arriving (invoices, contracts, leads, tickets) flowing into the eClips Workforce fleet, with two outputs — actions taken automatically, and a human triage inbox for exceptions — and a 'human governs / machine acts' label across the middle.