Actions & Agents (Layer E)

Layer E makes a packet executable: it defines what an AI agent should do with the knowledge. A packet with actions[] + agent_instructions (plus Layers B/C/D) is L3. action_type, trigger_type, and step_type are closed vocabularies — an unrecognized verb is one no runtime can execute.

Action

Field	Type	Notes
id	string	Required
action_type	enum	Required — see below
trigger	object	Required — { trigger_type, schedule?, filter? }
steps	ActionStep[]	Required, ≥ 1
name · description	string
entry_step_id	string	First step (defaults to steps[0])
inputs · outputs	object	Typed I/O declarations
enabled	boolean	Default true

Action types

research · analyze · synthesize · export · notify · enrich · validate ·
publish · sync · monitor · generate · classify · embed · index

Trigger types

manual · on_source_added · on_analysis_complete · on_export ·
scheduled · on_kb_change · webhook · on_gap_identified

A scheduled trigger carries a cron schedule; any trigger may carry a filter (DirectiveCondition).

ActionStep

Field	Type	Notes
step_id	string	Required
step_type	enum	Required — see below
config	object	Step-type-specific (prompt, tool_name, url, condition_expr, loop_over, …)
output_var	string	Variable to store output in (readable by later steps)
retry	object	{ max_attempts, backoff_ms }
timeout_ms	integer
next_step_id	string

Step types

llm_call · tool_call · http_request · condition · loop ·
parallel · emit_event · wait · transform · human_review

Agent instructions

agent_instructions is the structured brief for the consuming LLM:

Field	Notes
persona	The agent's role
system_prompt	Full prompt; {topic}, {domain}, {source_count} are template variables
goal	One-sentence objective
constraints	Hard rules the agent must never violate (e.g. "Never present contested findings as validated")
allowed_tools	Permitted tool names
output_format	Expected output shape
temperature (0–2) · max_tokens · model_hints[]	Generation hints (e.g. ["claude-opus-4-8"])

→ Next: Knowledge Graph · Directives (Layer D)