Agent_X is a layered framework for building agent systems safely. L0 stays protected. L1 governs implementation through root documents, FIC units, validation, evidence, and review. L2 proposes profiles and specifications before they become implementation work.
Many agent projects grow through ad hoc prompts, tools, and patches. Agent_X starts from explicit boundaries: what may change, who governs it, what evidence is required, and when work must be blocked.
Agent_X keeps the protected seed stable. Larger changes should happen through governed documents, FIC units, validation, and review.
If a change does not have the right goal, scope, ownership, or validation path, the correct result is blocked rather than guessed implementation.
Agent_X treats tests, traces, validation output, review records, and completion evidence as first-class project artifacts.
Agent_X separates the protected seed, implementation governance, and profile/specification work so future capability can grow without blurring authority.
The small stable base. It should remain protected from uncontrolled self-modification and broad runtime changes.
The layer where implementation is governed through root goals, whole-system pseudocode, shared types, FIC units, validation, and review.
The layer where specialization profiles, blueprints, integration specs, and evaluation specs are proposed before implementation.
A change should produce enough evidence for its scope: inspection, tests, validation output, traceability, and review records.
Agent_X turns broad ideas into governed implementation work. Root documents define the goal, whole-system pseudocode defines the flow, shared types define contracts, and unit FICs define bounded changes.
Stable public entrypoint for booting and running governed seed turns.
Selects the planned action while remaining constrained by contracts and boundaries.
Allows, denies, or requires approval before anything reaches execution.
Memory, evaluation, trace, and checkpoint ports preserve replayable proof of execution.
Agent_X can support different agent directions, but each direction should begin as a profile or specification before it becomes implementation.
Plan work, track tasks, manage files, organize priorities, and prepare bounded work packets.
Coordinate tools, agents, workflows, validations, and handoffs while staying inside governance boundaries.
Act as a higher-level decision layer, but still depend on explicit authority, review gates, and traceability.
Coordinate multiple specialized agents. This should come after simpler profiles because it has the highest governance risk.
Convert approved FIC units into code changes, tests, evidence, and completion records.
Explore sources, summarize findings, and produce evidence-backed research outputs without bypassing review.
Organizer, orchestrator, master-agent, and swarm examples should live in L2 as profiles or examples until one is selected for governed implementation.
Start with the L1 governance documents, then review the L2 profiles. Treat profiles as proposals until L1 converts them into governed implementation work.
L1/docs/00_system_goal.md
L1/docs/01_whole_system_pseudocode.md
L1/docs/02_shared_types_and_interfaces.md
L1/docs/FIC_GOVERNANCE_INDEX.md
L1/fic/
L2/docs/
L2/profiles/Do not jump straight from an idea to runtime code. Define the profile, derive the FIC unit, implement the smallest allowed change, validate it, and review before acceptance.
L2/profiles/
L2/blueprints/
L2/evaluation_specs/L1/docs/00_system_goal.md
L1/docs/01_whole_system_pseudocode.md
L1/docs/02_shared_types_and_interfaces.md
L1/fic/<unit>.mdimplementation files
tests
validation evidence
review packetUse Agent_X to describe, constrain, validate, and review agent evolution before it becomes accepted implementation.
Have questions about Agent_X, governed agent evolution, or Astrocytech products? Contact the team.