04 — AGENTS

Agents are assembled at runtime.

Each agent comes together at the moment of invocation from three streams — who the agent is, what it's been asked to do, and what it needs to know. The result is consistency where it matters, and adaptation where it counts.

WHAT IT DOES

Same role, different invocation, different output.

An Altabric agent is not a fixed entity sitting in a queue waiting for work. It comes together for one invocation, does the job, produces a deliverable, and is replaced by a fresh composition the next time round.

Three streams feed every invocation. The first is identity — the role prompt, the project-specific skills, the organisational grounding. This stream is stable: the same Data Architect serves a hundred invocations across the engagement without drift. The second is the task — a specific brief handed in by the workflow at this exact step. The third is the context — the documents and prior outputs the agent needs to do this particular task well, assembled at runtime by the engine.

The same role, given different tasks and different contexts, produces different outputs. That's not inconsistency — it's the agent doing what it should: applying stable expertise to a specific brief with specific material.

THE THREE STREAMS

Three streams converge. The agent emerges.

Identity is the slowest-changing stream — picked once per role, reused across the engagement. The task changes every invocation. The context is built fresh per work item. The agent itself doesn't persist.

IDENTITY · WHO YOU ARE TASK · WHAT TO DO CONTEXT · WHAT YOU KNOW Role prompt Job, competence, behavioural guidance. Who the agent is as a working professional. Skills Project-specific expertise. What the agent knows that is specific to this engagement. Vision and mission Organisational grounding — who the agent works for, purpose, values. Task template The specific brief handed in by the workflow at this exact step. Your documents From the library you've uploaded. The substrate the engagement runs on. Engine-assembled context Parent deliverable, linked items, prior outputs — at the compression the task needs. Different tasks for the same role get different context. The agent composed at invocation A deliverable

Identity is reused across invocations. Task and context are fresh every time. The agent itself is composed when the work begins, and a new one is composed when the next work begins.

THE ROLE LIBRARY

A defined set of roles. Each picked into workflows where they belong.

A working selection of the agent roles Altabric runs today. Each role has a remit, a default skill set, and a model tier chosen for the kind of work it does. Heavy reasoning runs on Opus, day-to-day production on Sonnet, lightweight summarisation on Haiku — keeping cost proportional to value.

Chief Technology Officer
Opus
Technical architecture, design integrity, cross-cutting technology decisions.
Chief of Staff
Opus
Programme coherence, sequencing, escalations, leadership support.
Chief Strategy Officer
Opus
Strategic intent, alignment of work against programme goals.
Data Architect
Opus
Data model integrity, schema decisions, integration patterns.
Product Owner
Sonnet
Business coverage, acceptance criteria, value alignment.
WBS Analyst
Sonnet
Decomposition review — are the children of this work item sized correctly?
Test Manager
Sonnet
UAT strategy, test plan authorship, defect process.
Test Analyst
Sonnet
Test pack production, scripts, test data, traceability.
Scrum Master
Sonnet
Sprint shaping, blockers, work-item sequencing within a workstream.
Change Manager
Sonnet
Stakeholder analysis, communications, training pack design.
Summariser
Haiku
Document summarisation, classification, lightweight ingestion tasks.
Human user
Human
Final authority. Approves, sends back, rejects. Where it matters.

The full library has more roles than shown. Each engagement gets a configured set matched to the delivery vocabulary you actually use.

EXAMPLE

One role. Two invocations. Two outputs.

A DATA ARCHITECT, ACROSS TWO WORK ITEMS

The role stays the same. The brief and the context don't.

A Data Architect is invoked to review the Epic-level Solution Design for a customer migration workstream. Identity stream: the Data Architect role prompt, the project-specific data-modelling skills, the organisational vision. Task stream: review this Solution Design for data-architecture coherence with the rest of the programme. Context stream: the Solution Design itself in full, the Theme PID at rich summary, sibling Epic designs at short summary.

Three days later, the same Data Architect is invoked again — this time on the Theme-tier integrating Solution Design. Identity stream: identical. Task stream: review the integrating design for data coherence across all Epic-tier designs. Context stream: the integrating design plus the linked Stories from across the Theme — different documents, different compressions, different reach.

Same role. Same expertise. Two clearly different outputs, because the brief and the material were different. That's the point of composition: the agent's expertise is reusable; the work it produces is task-specific.

TALK TO US

A pool of AI agents ready to accelerate your plans.

A demo runs against your delivery shape, not a generic one. We read every message. You will hear back within two working days.

Request a demo