Team lead
Prioritises, splits and coordinates. Assigns work, holds the office context together and decides whether the assignment is professionally fulfilled. Not a chat persona but a versioned, server-bound role.
How it works
Human-led.
Not boundlessly autonomous.
A good agent system does not prove itself by never stopping. It proves itself by stopping precisely when a release, the context or the evidence is missing. That is why the flow is built the way it is.
Role model
Three control roles, six specialist roles.
A complete office has nine agents. Three of them check and lead, six do the professional work. That separation is not organisational cosmetics — it is why a result is worth something in the end.
Security
Checks inputs, files, work paths and results against security, privacy and release rules. A negative or missing vote blocks the handover — without exception.
Research & Data
Sources, assesses and structures information. Checks sources, data quality, plausibility, contradictions and open evidence questions. Stops a convincing but poorly evidenced answer from passing as a result.
6 specialist roles
Tailored to your actual use — content, design, web, market analysis, sales, data analysis, customer care and more. Each role gets only the tools, data classes, budgets and execution rights its task requires.
From assignment to result
Twelve steps that build on each other.
- 01An authorised person describes goal, context, priority and expected result.
- 02The assignment is tied to a meeting, project, office and operation.
- 03Texts and files pass the security gate.
- 04Research & data assesses sources, data and open questions.
- 05The team lead turns this into bounded work packages.
- 06Specialist roles work their sub-tasks with released skills and tools.
- 07Peer reviews and iterations improve the result.
- 08Security, research & data and the team lead review the specific revision.
- 09Only on full release is a publication job created.
- 10The technical publisher provides an unaltered copy.
- 11You read the released file via your network drive.
- 12Corrections run as a new task and a new revision — a released state is never silently overwritten.
Security gate
Nothing reaches your agents unchecked.
Size, type and structure
Size, type, MIME and magic-byte checks plus archive and structure validation. What cannot be determined cleanly does not pass.
Malware and malicious code
Malware scan, checks for secrets, scripts, macros and links — before expert work, not after.
Prompt injection
Safe normalisation, content extraction and an explicit prompt-injection and policy assessment.
Four outcomes
RELEASE, MANUAL_REVIEW, REJECT or QUARANTINE. The architecture is fail-closed: uncertainty never widens a permission.
Content that cannot be checked, or that contradicts itself, is not waved through for convenience. That occasionally costs a follow-up question — and spares you the case where nobody can say where something came from.
Triple release
Three roles, one result, one revision.
Security
Is the output safe and policy-compliant?
Research & Data
Are data, sources and plausibility sufficient?
Teamlead
Does the revision fulfil the assignment and the quality standard?
Releases are bound to revision, operation and time. Revoked, expired or foreign-revision decisions must not be reused — otherwise the review would be a formality.