Agents with real authority, and real boundaries.
They plan, call your tools and finish multi-step work — inside the permissions, guardrails and budgets you set. Deployed on your own infrastructure.
Six decisions, not a prompt box.
Every agent in the platform is configured across the same six surfaces. Walk them in order and you have described precisely what the agent is, what it knows, and what it is permitted to do.
Identity — how it introduces itself
Name, avatar, description, and a domain template (general, creative, legal, medical, education) that seeds the agent's register. This is the part end-users actually meet.
Agents/Compliance OfficerPrivateIdentity
EAgent name
Compliance OfficerDescription
Expert compliance assistant for ISO 27001:2022 ISMS and SOC 2 management.Greeting message
I can help you manage your ISO 27001:2022 and SOC 2 compliance programme. Ask me about posture, gap analysis, policy drafting, or audit readiness.
Agent type
GeneralCreativeLegalMedicalEducationChanges are saved when you click Save.Save changesBehaviour — instructions and skills
The system prompt, the instruction modules layered on top of it, and the reusable skills the agent carries or loads on demand.
Agents/Compliance OfficerPrivateBehaviour
System prompt
## Working method
Prefer graph-aware tools over raw queries.
Cite the control or clause behind every answer.
Active instruction modules
Knowledge GuardrailGeneral HeaderTool Usage PolicySkills
Evidence assessmentAttachedPolicy exception reviewAttachedChanges are saved when you click Save.Save changesKnowledge & Privacy — what it may read
Which bases and datatables it can reach, whether it is confined to your files, and whether personal data is masked before it ever reaches a model.
Agents/Compliance OfficerPrivateKnowledge & Privacy
Knowledge base guardrail
Protected Mode · SecureBase & DataTable access
Compliance ManagementViewerControl registerViewerData protection
PII Masking · BetaWorld Knowledge · offChanges are saved when you click Save.Save changesTools & Integrations — what it may call
Scoped tool permissions, MCP servers, assigned workflows, and chat-platform bots. Permission is granted per scope, not per agent.
Agents/Compliance OfficerPrivateTools & Integrations
Tool permissions
base:*datatable:*data:*model:usecompliance:ReadmoreTool loading
Search FilesPreloadedRead File ContentPreloadedSearch Text in FilePreloadedIntegrations
MCP serversWorkflowsSlack botChanges are saved when you click Save.Save changesModels — and what happens when one is down
An ordered stack rather than a single choice. The default answers first; the rest are tried in order behind it.
Agents/Compliance OfficerPrivateModels
Model configuration
1Gemini 3.1 ProDEFAULT2GPT-5.23Gemma 4 (open weights)Models are tried in order. The default model is used for initial responses.
Temperature
Top P
Changes are saved when you click Save.Save changesLimits & Budgets — the ceiling it works inside
Step, tool-call and token caps, a soft timeout, and when context gets compacted. An agent that cannot run away is an agent you can leave running.
Agents/Compliance OfficerPrivateLimits & Budgets
Budget preset
StandardExtendedCustomEffective values
Max steps per turn20Max tool calls per turn25Max duplicate identical calls3Max resource scope per query3Max subagent depth3Soft timeout5 minutesChanges are saved when you click Save.Save changes
Agent settings: one surface at a time, in the order the console lists them
Some agents answer people. Some answer other agents.
An agent's mode decides who is allowed to call it and what it does with work it cannot finish alone. Pick a mode to see the shape it creates.
Standalone — Operates directly with end-users and tools.
In standalone mode a support engineer messages the runbook assistant directly, and the agent calls its own file tools. Nothing sits in between.
- thenEnd userSupport engineermessagesAgentRunbook assistant
- AgentRunbook assistantcallsToolfile:*
Mode is a per-agent setting, not a plan tier. The same agent can be promoted from standalone to hybrid the moment an orchestrator has a reason to call it.
Instructions that don't drift.
An agent's system prompt is the part everyone edits and nobody reviews. The platform layers it: your instructions sit on top of modules the workspace enforces, so the guardrails cannot be prompted away.
Knowledge Guardrail
Confines answers to your indexed files and requires a citation for each one.
General Header
Workspace-wide framing every agent inherits, so register stays consistent.
Tool Usage Policy
How and when to reach for a tool — and when to ask a human instead.
Active instruction modules are applied by the workspace, above whatever the agent's own prompt says.
# How to write a 10x skill
1. Define exact boundaries
Bad: “Be helpful.”
Good: “Never modify the production database unless explicitly requested.”
2. Specify the output format
Bad: “Give me the results.”
Good: “Output the results as a JSON array of objects.”
Write constraints, not encouragement.
Agents perform best when they know exactly what they may not do. Vague warmth produces vague behaviour; explicit boundaries and named output shapes produce the same answer twice.
- Define the exact output structure you expect
- State the negative constraints, not just the goals
- Give one worked example over three paragraphs of tone
Write the playbook once. Attach it everywhere.
A skill packages instructions, and optionally the tools they need, into something an agent can carry in every prompt or load on demand from its skill index. Author it once and every agent that should follow it, does.
A skill is three parts: a human name and an immutable slug, a when-to-use rule the model reads to decide whether the skill applies, and a Markdown body holding the method.
1. Author it once
Create a skill
Reusable instructions for agents.
Name
Slug
When should an agent load this skill?
Category
Security operationsActivated tools
# Incident response brief
## When to use
Use for security incidents, triage updates, or requests for an executive brief.
## Working method
1. Establish scope and severity from the available evidence.
2. State facts separately from assumptions.
3. Surface the next accountable action and owner.
Approved tools
2. Attach it per agent
Skills
Attach reusable guidance this agent can load as requests require it.
Give every agent a focused skill set.
Attach published skills from the shared library to compose a specialist without rewriting its system instructions — and without every agent carrying every instruction.
- Attach a published skill to one agent or a repeatable role
- Update shared guidance once, not per agent
- Pair expertise with only the tools it is allowed to use
Agents inherit the workspace's default language and answer in 120+ others when the conversation calls for it.
Capability is granted by scope, and paid for in context.
Every tool an agent can call is a permission you granted and a slice of its context window. The platform separates the two decisions: what it may do, and what it should carry with it.
Tool loading
Preloaded tools are always available to the agent. Discoverable tools are found on-demand via search, reducing initial context size.
DATA
FILES
MODEL
Context window at start of chat
7 of 12 tools preloaded
Every preloaded tool spends context before the user has said anything. Discoverable tools cost nothing until they are needed.
Enabled scopes
base:*datatable:*data:*model:usecompliance:Readfile:*folder:*Wildcards grant a family; an explicit verb like compliance:Read grants exactly one. Anything not listed is denied — there is no implicit access.
MCP integrations
Reach external services through MCP servers you connect — 170+ available, plus any server you host yourself.
Workflow assignments
Assign workflows users can trigger from this agent's chat, surfaced as a Run a workflow button.
Chat platform bots
Connect the same agent to Slack so it answers in the channel where the work already happens.
Rich output
Let the agent render inline charts, KPI cards and tables in chat. Off by default — when disabled, no extra tool or instructions load, so the agent stays lean.
The tenth agent is a configuration, not a project.
Every agent here is the same six decisions, drawn from one skill library and one set of scopes. The second one starts from a working configuration, not a blank page. The tenth can be the one the ninth calls. Work that could never justify a system of its own becomes work you can hand over.
Guidance you write once is carried by every agent that should follow it, and you change it in one place.
An agent that cannot run away is an agent you can leave running.
Work that runs while nobody is watching still ends at a person, because the orchestrator holds the result until someone approves it.
An agent that cannot invent is an agent you can audit.
Grounding is not a prompt instruction here — it is a mode. Protected Mode confines the agent to what you gave it and makes every answer carry its source.
Protected Mode
Secure- Only responds with information from uploaded files
- Provides mandatory source citations
- Ensures accurate, verifiable responses
- File and folder tools are automatically enabled
Allow Web Search Citations
An explicit opt-in. When enabled the agent may search the web and show web citations alongside your file references — never instead of them.
Base & DataTable access
Grant an agent structured data the same way you grant a person: per resource, with a role.
Compliance Management
Base · All datatables
Control register
DataTable · Read only
The agent can use tools on these resources, but they never appear in the user's sidebar — access to an agent is not access to your data.
PII Masking
BetaMask sensitive values — tax IDs, government numbers and other personal data — before anything is sent to the model.
World Knowledge
Off by defaultWhen enabled, the agent may use the model's own knowledge or visual judgment, but only for knowledge-graph tasks you annotate, and always with a caveat shown to the user.
It is layered on top of the knowledge guardrail. It never turns the guardrail off.
- Runs in your own infrastructure
- No data leaves for inference
- Air-gapped deployment supported
- Full audit trail per answer
Grounded, masked, cited — and none of it leaves your network.
One agent, an ordered stack of models.
Choosing a model is a decision you will want to change. So it is a list, not a field — and the ceilings the agent works inside are set in the same place.
Model configuration
- 1DEFAULT
Gemini 3.1 Pro
Highest reasoning quality
- 2
GPT-5.2
Balanced quality and speed
- 3
Gemma 4 (open weights)
Runs local — the air-gapped floor
Models are tried in order. The default model is used for initial responses.
Sampling
Advanced settings
Temperature
Controls randomness. Lower is more deterministic.
Top P
Controls diversity via nucleus sampling.
Both are per-agent, and both are the reason a compliance agent and a drafting agent should never share a configuration.
Limits & budgets
Budget here means work, not money: how many steps, tool calls and tokens one chat turn may consume before the agent has to stop and report back.
Standard — Chat, Q&A and single-document tasks. Inherits the tenant defaults — right for most agents.
- Max steps per turn
- 20
- Max tool calls per turn
- 25
- Max duplicate identical calls
- 3
- Max resource scope per query
- 3
- Max subagent depth
- 3
- Soft timeout
- 5 minutes
- Compaction buffer
- 20,000 tokens
- Auto-compact at
- 85% of context
- Per-chat token cap
- Set per workspace
Every value is a per-agent override, not a product ceiling, so a drafting agent and a compliance agent never share a budget.
The agents we already run.
The six surfaces above are not a spec sheet. They are how we configured these, and each one runs in production on the customer's own infrastructure.
Medical OCR & document intelligence
Clinical documents, prescriptions, lab reports and insurance forms turned into structured, coded data — with medical named-entity recognition and EHR/FHIR mapping behind it.
- 99.2% extraction accuracy on clinical documents
- Medical NER and entity linking, not generic OCR
- EHR / FHIR mapping and pre-built connectors
- HIPAA-aligned, deployed inside the hospital network
Lab report · PDF
Structured · FHIR
- Patient
- Masked
- Diagnosis code
- ICD-10 · E11.9
- Specimen date
- 2026-03-11
- Result
- 6.4 mmol/L
Compliance Officer
Runs an ISO 27001:2022 ISMS and SOC 2 programme: gap assessments, risk analysis, policy drafting, evidence review, audit preparation. It reads a typed compliance graph rather than guessing from documents.
Media & content intelligence
Transcription, translation and drafting across 120+ languages with speaker diarization — built for newsrooms and broadcast feeds where speed to publish is the metric.
Build your own
Nothing above used a private API. Same six surfaces, same guardrails, same on-premise deployment — pointed at your domain instead of ours.
Every one of these started as an empty agent and six decisions. That is the whole product.
How an agent actually ships.
Creating the agent is the quick part. What makes it safe to hand to three hundred people is everything after it.
- 01
Create
Name it, give it a description and a domain template, and point it at a filestore. It exists, and it can do nothing yet.
- 02
Configure
Work the six surfaces: instructions and skills, what it may read, which scopes it may call, its model stack, its budget.
- 03
Grant access
A private agent is invisible until a group is granted it. Access is a deliberate act, per group — never a default.
- 04
Share as a package
Bundle the agent with the bases, skills and tools it depends on, so the same working configuration lands in another workspace intact.
- 05
Watch it, then tighten it
Usage, latency and cost per agent come back as telemetry. Raise the budget where it earns it and cut the tools it never called.
See Exemplary AI in action
Book a demo and we'll show you purpose-built agents grounded in your own knowledge — deployed on your infrastructure.