Pipeline with Memory, not Swarm

Deterministic agents for auditable work.

AMS turns AI automation into a traceable pipeline. Every action is logged, every decision is chained, and every change can be verified or rolled back.

GitHub repo Quick start

LiveMCP tool count via unified_help

17+SQLite tables active

LiveRoadmap count via roadmap_list

Phase 1+2+3Core complete

What AMS is

AMS is an Agent Management System that unifies GSD, roadmaps, and unified MCP into one deterministic pipeline with RBAC/ACL, skill and agent spawning, graph-aware analysis, create/register/init/help operations, memory frames, cryptographic audit chains, Merkle proofs, and rollback-safe checkpoints.

Inputs: intent, project context, constraints

Process: deterministic pipeline with memory

Outputs: tasks, plan, proofs, verified changes

Agent pipeline

INIT -> GATHER -> ANALYZE -> PLAN -> [DRY_RUN | APPLY] -> VERIFY -> DONE

Phase intent

INIT: register or load project, create Git checkpoint

GATHER: collect context, repo facts, roadmaps

ANALYZE: detect gaps, risks, or learning targets

PLAN: generate prioritized tasks (Eisenhower matrix)

DRY_RUN/APPLY: preview or execute changes

VERIFY: confirm results and tests

DONE: finalize session, attest and archive

Why pipeline over swarm

Swarm	AMS pipeline
Distributed state	Single source of truth
Race conditions	Sequential steps
Unclear blame	Action-level accountability
Hard rollback	Git checkpoints + rollback

Artifacts and memory

Persistent memory

SQLite stores sessions, tasks, contexts, audit entries

Hash chain records every tool action

Merkle tree finalizes proofs for verification

What you get back

Prioritized task matrix (Q1 to Q4)

Session audit trail with timestamps

Git checkpoint hashes for rollback

Exportable proofs for compliance

Safety and control

Built-in safety gates

Dry run mode for safe planning

Max files limit to bound blast radius

Auto rollback on failure

Rate limiter and circuit breaker

Access control

Role-based access via AMS_USER_ID

Owner, admin, member, viewer roles

Default deny for unknown tools

Per-project isolation

Project status

Phase 1 Foundation complete: audit system, hash chain, merkle proofs, Git checkpoints, SQLite persistence.

Phase 2 Autonomous mode complete: intent detection, task planning, safe apply/rollback, verification.

Phase 3 Complete: hybrid RAG, strict quality mode, tool profiles, and AMS-QR structured memory.

Quick start

MCP client config

{ "mcpServers": { "ams": { "command": "node", "args": ["/path/to/AMS/projects/unified-mcp/src/server.js"], "env": { "AMS_ROOT": "/path/to/AMS" } } } }

Run the pipeline

ams_autonomous_run { "project_name": "my-project", "intent": "explore", "safety": { "dry_run": true } }