Reasoning engine

When a situation is detected, the reasoning engine builds an action plan. It does so by reading the wiki — not by improvising from pretrained weights, and not by querying a vector store in isolation.

Context assembly

Before reasoning begins, context is assembled from multiple sources:

• Entity properties and related entities — the situation’s anchoring entities plus graph-adjacent data.
• Activity timeline — 30 days of behavioural patterns from person and department pages.
• Communication context — pgvector retrieval of relevant email threads, Slack messages, and document excerpts.
• Cross-department signals — for external entities, signals from adjacent departments.
• Playbook and heuristics — situation-type page content including canonical sources and learned patterns.

Four-stage decomposition

A reasoning cycle decomposes into four stages with distinct cognitive shapes:

• Understand. Classify the task. Load the playbook. Assemble seed context. Lightweight.
• Investigate. Read evidence. Build hypothesis. Reduce uncertainty. Tool-heavy, retrieval-driven.
• Decide. Commit to a position. Generate a decision manifest — what claims will be made, what actions proposed. Few tool calls, high reasoning density.
• Produce. Write deliverables with manifest-driven retrieval. Each committed claim triggers a re-fetch of ground-truth context for citation fidelity. Verification before commit.

Between stages, a narrowing operation compresses prior-stage noise while preserving distilled understanding. When playbook confidence is high and complexity is low, the four stages merge into a single pass.

Single-pass vs. multi-agent

Small situations (under ~12K context tokens) run single-pass. Large situations activate the multi-agent path: three specialists — Financial, Communication, Process/Compliance — plus a coordinator that synthesises. This keeps per-call context bounded without losing the combined signal.

Reflection tools

The reasoner’s arsenal includes three classes of self-facing reflection tools:

• Curiosity tools widen consideration space (probe_adjacent, explore_implications). Active during Investigate.
• Skepticism tools challenge committed positions (steelman_opposite, find_disconfirming). Active during Decide.
• Verification tools check committed claims against ground truth (verify_citations, check_consistency). Active during Produce.

Each reflection invocation spawns a scoped sub-invocation with compressed context. The main reasoner evaluates the output and decides how to integrate. Reflection is required at the end of each project-deliverable Produce stage; elsewhere it is activation-gated but elective.

Model routing

Anthropic Opus 4.7 handles strategic reasoning — onboarding synthesis, deep investigations, situation reasoning, reflection tools for high-stakes tasks. Sonnet handles iterative investigation loops and standard reflection invocations. Haiku handles classification, extraction, activity detection, and post-resolution consolidation. Per-capability routing is a config change — model progress is free uplift.