How Decisional Agents Work

A technical explanation of Decisional's agent architecture—from natural language instructions to autonomous workflow execution.

Last updated: January 15, 2026

Agent Lifecycle

Agents progress through defined states as they move from creation to production:

DraftAgent instructions being edited. Not ready for execution.
VerifiedInstructions entered with trigger and integrations configured. Goal validated, workflow generated.
TestedAt least one test run completed successfully.
LiveAgent is running in production.
PausedAgent paused. No runs will execute until resumed.

Two-Tier Authoring and Execution

Decisional uses a two-tier system that separates workflow authoring from execution:

Tier 1: Authoring (Alakazam)

When you describe what you want done, an AI-powered authoring system interprets your natural language instructions and generates a structured workflow:

  • • Parses natural language instructions
  • • Generates workflow DAG (Directed Acyclic Graph)
  • • Tests workflow logic in sandbox environment
  • • Produces execution-ready workflow definition

Tier 2: Execution (Runner)

When the agent runs, a separate execution engine processes your data:

  • • Executes workflow nodes in dependency order
  • • Processes spreadsheet rows in parallel
  • • Handles retries and error recovery
  • • Stores results and artifacts
"This separation allows the AI to focus on understanding your intent during authoring, while deterministic execution ensures reliable, repeatable runs."

Row-Based Processing Model

Decisional's spreadsheet-native architecture processes data row by row:

  • Rows as independent documents: Each row represents a discrete task (an invoice to process, a quote to generate, a lead to research).
  • Sequential task execution: Tasks within a row execute in order, building on previous results.
  • Parallel row processing: Multiple rows can be processed simultaneously for throughput.
  • Field accumulation: As tasks complete, new fields (columns) are added to the row with results.

Runtime Reasoning

Unlike rule-based automation, Decisional agents reason at runtime:

Traditional Automation

  • • Pre-programmed conditions
  • • Fails on unexpected input
  • • Every edge case must be anticipated
  • • Rigid data format requirements

Decisional Agents

  • • Understands intent, not just rules
  • • Adapts to input variations
  • • Reasons through edge cases
  • • Handles diverse data formats

Example: When processing invoices, if one invoice has a different format, the agent reasons about where to find the relevant information rather than failing because fields aren't in expected positions.

Human-in-the-Loop Checkpoints

For decisions that require human judgment, agents can pause for approval:

  • Approval triggers: Defined conditions (amount thresholds, uncertain extractions, flagged items) pause execution.
  • Review interface: Users see what the agent proposes, with context from the workflow.
  • Approval actions: Approve, reject, or modify the agent's decision.
  • Audit trail: All approvals are permanently stored with full context for compliance.

Versioning Model

Decisional maintains version history for reproducibility and auditing:

Live Version

One per agent. Continuously updated with user changes. No version number—it's the current working version.

Run Snapshots

Created at run time. Immutable. Numbered sequentially (v1, v2, v3...). Each run is linked to a specific snapshot, ensuring you can always see exactly what configuration was used.

Change Detection

Content hashing prevents duplicate versions. A new snapshot is only created if the agent configuration has actually changed.

Workspace and Worksheet Model

How Decisional manages your data during execution:

  • Live worksheet: Remains untouched during normal operation. This is your source of truth.
  • Run locking: When a run starts, the worksheet is locked (read-only with "LOCKED" status) to prevent conflicts.
  • Agent workspace: Editable by the agent during execution. Results are written here.
  • User workspace: Read-only for the agent. Ensures your data isn't modified unexpectedly.

Agent Instructions

Instructions are the natural language document that tells the agent what to do:

  • Plain English: Write instructions like you'd brief a colleague.
  • Snippets: User-specific instructions (custom prompts, business rules) that can be reused across agents.
  • Triggers: Conditions that start the agent (schedule, manual, webhook).
  • Integrations: Connected tools the agent can use (email, browser, APIs).

Technical Capabilities

Document formats

Parallel

Row processing

Full

Version history

100%

Audit trail

Learn More

See agents in action

Experience AI-powered workflow automation.

Visit main site