The service every action passes through
One relational model for tasks, runs, tool calls, artifacts, approvals and audit entries. Risk classification, approval requests and the append only ledger all live here, not in the agents that call it.
The spine comes first. Everything else runs on it.
Most AI stacks bolt governance on at the end, where it becomes a policy document nobody can enforce. Arakno starts with the record layer, then builds knowledge, research, orchestration and delivery on top of it.
What the business knows, with sources
Memory holds who the customer is and how you work. The knowledge fabric holds reusable company knowledge as versioned pages with immutable snapshots, so every answer can point at where it came from.
The outside world, on a budget
Web, feed, code and video adapters behind one gateway with request guards, an explicit query and time budget, and scoring that rewards independent corroboration rather than repetition.
Specialists, not one generalist
A parent run can delegate to sub agents with isolated context, cap how many run at once, and pick a different model per job. Results come back to the parent with the trace intact.
Durable work, not just messages
Documents, reports and exports are indexed, versioned and searchable, with safe preview and download. Work an agent produces survives the conversation that produced it.
Goals measured by events, not vibes
Goals decompose into tasks and report progress counted from real events. Overnight, the system reviews its own evidence and proposes improvements for a human to accept or reject.
Seven levels of autonomy, and you choose where to stop.
The same platform runs at every level. Most customers start at the second and move up when the evidence justifies it.
Assisted
- Level 1, single requestsYou ask, the agent answers, the exchange is recorded. Useful, and about ten percent of the value.
- Level 2, memory and identityThe agent remembers the business across sessions and answers in your voice, with your constraints.
Directed
- Level 3, commands that multiply outputBackground work, mid flight redirection, queued follow ups and model switching inside one session.
- Level 4, skills and connected systemsNamed skills, each on the model that suits it, reading and writing through your channels, calendar, CRM and files.
Autonomous
- Level 5, orchestrationParallel sub agents with isolated context and an explicit budget, evaluated before the result is accepted.
- Level 6 and 7, standing workPersistent goals, scheduled jobs and overnight review. The shift is from asking and answering to working and reviewing.
The right model per job, with the cost attached.
Paying a premium model to reformat a spreadsheet is the most common way an AI programme quietly loses its budget. Arakno routes each skill to a model chosen for that skill, and records what the run cost.
- Research and high volume extraction on a cheap, fast model.
- Drafting and customer facing writing on a model chosen for tone.
- Code review, reasoning and adversarial checks on a frontier model.
- A routing panel where you can see, and change, which model serves which skill.
SkillModel tierWhyClass
lead researchfasthigh volumeR1
reply draftingbalancedtone mattersR2
refund decisionfrontiermoney movesR4
nightly reviewfrontierjudges itselfR1
Example routing. Yours is set during onboarding and can be changed without a redeploy.
Where it runs.
Deployment shape
- One isolated instance per customer, not a shared tenant
- Internal services bound to loopback, never exposed to the internet
- Operator surface behind mutual TLS and authentication
- Hosted by us, or in your own cloud account when policy requires it
Operational discipline
- Health checks on every service, with restart counts visible
- Changes staged behind a flag and reversible in one step
- Rollback snapshot kept before each deployment
- Credentials registered by reference and rotated on a schedule
See it running against your own workflow.
We will take one process you actually run, and show you the trace it produces end to end.