Product

A control plane for your agent fleet.

The fleet is an inventory with lenses, not a wizard. Every view is a projection of the same inventory — with owners, budgets and audit from day one.

01·Mental model

Three concepts hold everything up.

The agent as an operating unit

Unit

The core entity. An autonomous agent with a lifecycle state, a human owner and a primary group. Not a service — a position that's accountable.

The contract between business and implementation

Role

Mission, explicit limits (what it does NOT do), required access, autonomy, KPIs and budget. Lives in Kordana even when the build lives in another framework.

The org chart for the fleet

Group

A hierarchy with business meaning — domain, customer, process — with a human owner at every node. Budgets and policies inherit downward.

02·Five operating surfaces

The screen that passes the meeting-screenshot test.

Designed as case management, not chat. The retention engine lives in the inbox; the business conversation lives in the control room.

128 agents active

Fleet spend (today)

$13,520

/ $18,900 budget

Open escalations

3 waiting on decision

Fleet SLA

97.2%

target 96%

Units — Customer support

24 active

Unit	Role	Owner	Spend	KPI	Status
UNIT-012	Tier 1	M. Salas	$212	98.4%	Operating
UNIT-019	Returns	J. Ortega	$174	91.2%	Supervised
UNIT-031	Billing	L. Mendoza	$156	99.1%	Operating

Fleet control room

Map by groups with an aggregated status, work in progress, spend for the period and active alerts. Drill down domain → group → unit.

Escalation inbox

Cases queue ordered by priority and SLA, with full context. Three outcomes: resolve, return with instructions, take the case.

Unit page

State, recent work, live KPIs, spend vs. budget, version history and operational controls. The full record.

Incident panel

Timeline in business language, affected units, containment actions and comms. "Collections-LATAM degraded since 09:40".

Fleet payroll

Spend by domain, group, unit and period. Projected close, budget vs. actual, threshold alerts. Financial language.

03·Workflow

Six verbs that govern the fleet, in the same loop.

Everything that happens in Kordana falls into one of six verbs. Every step leaves a trace in the inventory and the audit.

Step 01Live roster
Inventory. Bring what already runs onto the map.
Every agent enters with an owner, group and mission — L0 manual through L3 deep control. The inventory comes before the conversation.
Step 02Role certified
Certify. Turn the role into a contract.
Mission, limits, access, KPIs and budget are signed off. An implementation ships to production only if it meets the role.
Step 03Aggregated status
Operate. See health, work and spend live.
The Control Room aggregates status, throughput and cost by group. The business conversation runs through that screen.
Step 04SLA met
Escalate. Decisions with context, not alerts.
The inbox routes cases with SLA to the accountable human. Three outcomes: resolve, return with instructions or take the case.
Step 05Signed action
Control. Intervene without redeploys.
Stop, freeze, degraded mode, inline approval or live limits. Every lever is signed into the audit.
Step 06Next version
Review. Close the loop against the contract.
Scorecards by role, findings and experiments that translate into the next version of the unit.

04·Real operational control

Five levers to intervene without redeploys.

Stop / Play

Stop or resume a unit in one click.

Fleet freeze

Emergency button by group or entire fleet.

Degraded mode

The unit runs with reduced autonomy: every action goes through approval.

Inline approval

Human sign-off before sensitive actions, with SLA.

Live limits

Adjust budgets, hours and action caps without a redeploy.

Every control action is written to an immutable audit: actor, object, timestamp and justification.

05·Metrics that govern

From throughput to spend, with the autonomy rate as the flagship metric.

Autonomy rate

Flagship

% of work completed without human intervention

Operational availability

per group, with business thresholds

Throughput vs. expected

compliance against the role contract

Resolution time

of escalated cases vs. SLA

Spend vs. budget

with alerts to the owner chain

Incident MTTR

on formal incidents

06·Intelligent runbook

Fixed graph, LLM inside.

Every Unit runs on a single standard LangGraph template. The LLM plans sub-goals, drafts and decides tool calls inside the box; it doesn't reorder or create nodes. What is not allowed does not exist in the topology.

receive

plan

act

verify

close / escalate

One deterministic graph for the whole fleet — no ad-hoc scripts per agent.

07·Composer + Role Library

Design roles, not scripts.

A Role is the position —mission, limits, access, KPIs, budget. A Unit is the agent that fills it. Both versioned and audited.

Composer

Design Roles as contracts: what it does NOT do, what it needs, how it's measured. Executable, signable output.

Role Library

Per-tenant catalog plus a global template library. Publish, version, promote with full lineage.

08·Model gateway & BYOAI

Bring your keys. Keep governance.

The gateway routes by quality tier and bucket (A/B/C), resolves encrypted BYOAI credentials just in time, enforces budget and emits token accounting. It never accepts a client-supplied model name: it routes by tier.

OpenAIAnthropicGoogleAzureBedrockVertexSelf-hosted

Quality-tier routing
A/B/C buckets with deny-list
Per-workspace encrypted credentials
Real-time token accounting

09·Tools, knowledge, freshness

Governed action and context planes.

Composio tools

Curated toolkits over standard MCP. Per-action allowlists, encrypted non-human identities per connection.

MCP knowledge

Sources by reference, never content. Kordana orchestrates access; the data never leaves its origin.

Corpus freshness

Per-source freshness SLA. Alerts when knowledge ages and answers begin to drift.

10·Quality flywheel

Escalations don't get lost — they become evidence.

Every time an agent escalates to a human, the answer is captured as typed, immutable structured feedback. It feeds the eval bank, human-signed certification and deterministic scorecards that gate the next version.

01
Escalation
The agent asks for help with context and pending decision.
02
Structured feedback
The human answers in types, not free text.
03
Eval bank
Versioned, immutable case with lineage.
04
Signed certification
A human promotes the version with audit.
05
Deterministic scorecard
Kordana computes the verdict — never an LLM.

11·Work windows & A/B experimentation

Work windows

Time-of-day policies per Role or Group. Shifts, scheduled freezes, maintenance exclusions and off-hours regimes.

A/B experimentation

Two variants of a Unit run in parallel. The winner is declared by deterministic verdict over scorecards — no LLM or human bias in the count.

How does your fleet reach Kordana?

Six phases, one inventory. Start by inventorying what already runs.

See how it works Start free