In Development

Drafts

These pages describe planned or partially implemented capabilities. They are collected here to avoid implying they are available today.

Automated Mastery (Draft)

Automated Mastery is the mechanism through which Squad progressively shifts from expensive, deliberate reasoning to fast, automatic execution. For the design philosophy and neuroscience inspiration behind this progression, see the sections on System 1 / System 2 throughout the docs. This section covers the practical mechanics.

Progressive Hardening

Workflows get faster and cheaper over time without sacrificing quality:

1. Novel task: Full System 2 reasoning. AIM plans from scratch, executes step by step, reviews each result. Expensive but necessary.

2. Familiar pattern: Template matching. AIM recognises the pattern from past traces and uses a proven plan as a starting point. Faster and more reliable.

3. Approved workflow: System 1 execution. The workflow runs as a compiled workflow: deterministic tool calls and constrained model operations. Fast, cheap, predictable.

Individual steps within a workflow follow the same trajectory. A step that initially requires a full agent reasoning loop can be demoted to a constrained model call (same output, less cost), and eventually to a pure deterministic tool (no model needed at all).

Workflow Hardening

When the system successfully executes a multi-step workflow, the full execution plan can be saved as a reusable workflow template. An administrator reviews the plan and its results; once approved, the workflow is stored in procedural memory and future matching requests execute the template directly: skipping expensive novel reasoning entirely.

Reuse counts and outcomes are tracked continuously, and workflows can be revoked at any time if they produce poor results.

Knowledge Crystallisation

While query approval and workflow hardening are human-driven, SOMA’s crystallisation process is designed to run in the background: discovering patterns across accumulated episodes and strengthening the knowledge graph. The crystallise node exists in the AIM graph, but the full background consolidation process is in development.

Crystallisation handles:

• Pattern extraction: Identifying recurring entity patterns and promoting them to typed categories
• Co-occurrence promotion: When entities frequently appear together across episodes, their implicit relationship is promoted to an explicit edge in the graph
• Confidence calibration: Adjusting confidence scores on entities and relationships based on accumulated evidence

Note

Crystallisation improves retrieval quality and entity resolution over time, but it does not create new workflows or change how the system acts. Behavioural changes (new workflows, new tools) always require human approval.

How the Platform Learns (Draft)

Squad is designed to get better with use. Every query, every correction, and every approval feeds back into the system’s knowledge and capabilities. But this learning is not autonomous: it happens under human control, at human-governed pace, with full visibility into what the system has learned and how it uses that knowledge.

The Mastery Cycle

When you first learn to drive, every action requires conscious thought: checking mirrors, coordinating pedals, judging distances. With practice, these sequences become automatic. You don’t think about how to drive; you just drive.

Squad follows the same progression: try, learn, get taught, remember. When it encounters a novel task, it reasons from first principles: planning, executing, reviewing, and recovering from failures. Every attempt is recorded as a reasoning trace. Domain experts review what worked and approve successful patterns as reusable workflows. The next time a similar task arrives, Squad executes the proven workflow directly: no deliberation needed.

The Learning Loop

Squad’s learning operates across three timescales, each with different human oversight requirements:

Timescale	What Learns	How It’s Governed
Per-session	Working memory, conversation context, active reasoning	Automatic: scoped to the current session, discarded after
Across sessions	Semantic knowledge: entities, relationships, patterns crystallised from repeated observations	Background process with human review of promoted knowledge
Slow, deliberate	Procedural memory: approved workflows, query templates, tool definitions	Requires explicit human approval before anything enters the system’s repertoire

The key design principle is that the most impactful forms of learning are the slowest. A workflow that executes automatically across the organisation requires the most scrutiny before it’s promoted: and the system enforces that.

What Makes This Different

Most AI systems are stateless: every query is processed from scratch with the same cost and latency. Squad’s mastery model means:

• Cost decreases with maturity. Early interactions are expensive (deliberate reasoning). Mature workflows are cheap (compiled workflows). The more you use Squad, the less each query costs.
• Reliability increases with use. Proven workflows have been validated by humans and tested through repeated execution. They’re more reliable than novel reasoning.
• Knowledge is organisational, not generic. The workflows Squad learns are specific to your data, your domain, and your workflows. This is institutional knowledge encoded as executable automation.
• Humans stay in the loop. Nothing enters procedural memory without human approval. The system proposes: domain experts decide what becomes permanent.