Agentic Project Management: Hierarchical Tasks for AI Workflows

Running AI agents on complex projects without structure is chaos. Hierarchical Tasks (htasks) bring project management primitives to multi-agent workflows: 4-level hierarchies, policy-driven governance, blocking propagation, and full audit trails.

Key Takeaways

4-level task hierarchy — Initiative → Feature → Workstream → Task
Automatic blocking propagation — blocked child → blocked parent
Policy-driven closure — require evidence, enforce waivers
Full observability — metrics, audit trails, checkpoint deltas

The Problem: Flat Tasks Don't Scale

Most AI agent frameworks use flat task queues: create a task, claim it, complete it. This works for simple workflows, but breaks down when you need:

Multiple agents working on related tasks
Dependent work that blocks on upstream delivery
Evidence that a task was actually completed correctly
Audit trails for compliance or debugging
Structured escalation when things go wrong

Real software projects have structure: initiatives break into features, features decompose into workstreams, workstreams contain atomic tasks. Flat queues lose this hierarchy — and with it, the ability to track progress or propagate status.

Hierarchical Task Architecture

Snipara's rlm_htask_* tools implement a 4-level hierarchy designed for real project management:

N0: INITIATIVE     "Q1 Platform Modernization"

│

     ├── N1: FEATURE     "User Authentication"

     │        │

     │        ├── N2: API          ← Workstream

     │        │    ├── N3: TASK    "Implement JWT endpoint"

     │        │    └── N3: TASK    "Add refresh tokens"

     │        │

     │        ├── N2: FRONTEND     ← Workstream

     │        │    └── N3: TASK    "Build login form"

     │        │

     │        └── N2: QA           ← Workstream

     │             └── N3: TASK    "Write E2E tests"

Level	Name	Description	Typical Owner
N0	Initiative	Strategic goal spanning multiple features	Product lead / Architect
N1	Feature	End-user deliverable	Feature lead
N2	Workstream	Parallel work track (API, Frontend, QA)	Tech lead
N3	Task	Atomic unit of work	Individual agent

Creating Tasks: From Initiative to Atomic Work

Quick Feature Scaffolding

The rlm_htask_create_feature tool creates a Feature (N1) with standard workstreams in a single call:

rlm_htask_create_feature(

  swarm_id="swarm_abc",

  title="User Authentication",

  owner="alice",

  workstreams=["API", "FRONTEND", "QA"]

# Creates N1 + 3 N2 workstreams automatically

Detailed Task Creation

For atomic tasks (N3), use rlm_htask_create with acceptance criteria and evidence requirements:

rlm_htask_create(

  swarm_id="swarm_abc",

  level="N3_TASK",

  parent_id="htask_ws_api",

  title="Implement JWT generation",

  owner="agent_backend",

  priority="P1",

  acceptance_criteria=[

    {"id": "ac1", "text": "JWT signed with RS256"},

    {"id": "ac2", "text": "Tokens expire after 1h"}

],

  evidence_required=[

    {"type": "test", "description": "Unit tests passing"}

Blocking Propagation: Status Flows Upward

When a task is blocked, the status automatically propagates up the hierarchy:

N1: Feature     → BLOCKED (child blocked)

    N2: API     → BLOCKED (child blocked)

        N3: Task → BLOCKED ← Original block

This means:

Project leads see blockers at the feature level without drilling down
Metrics reflect actual project health
Escalation is automatic when blockers affect upstream deliverables

Block with Context

rlm_htask_block(

  swarm_id="swarm_abc",

  task_id="htask_001",

  blocker_type="DEPENDENCY",

  blocker_reason="Waiting for auth API from team-backend",

  blocked_by_task_id="htask_auth_api",

  eta_recovery="2026-03-21T10:00:00Z",

  escalation_to="tech-lead@company.com"

Blocker types: TECH, DEPENDENCY, ACCESS, PRODUCT, INFRA, SECURITY

Policy Engine: Configurable Governance

Each swarm has a policy that governs task behavior:

rlm_htask_policy_get(swarm_id="swarm_abc")

# Returns:

  "closurePolicy": "STRICT_ALL_CHILDREN",

  "requireEvidenceOnComplete": true,

  "failedIsBlockingDefault": true,

  "allowParentCloseWithWaiver": true

Policy	Effect
`closurePolicy: STRICT`	Parent can only close when ALL children complete
`requireEvidenceOnComplete`	N3 tasks MUST provide evidence to close
`allowParentCloseWithWaiver`	Allow closing with documented exceptions

Evidence & Closure: Proving Work is Done

Complete with Evidence

rlm_htask_complete(

  swarm_id="swarm_abc",

  task_id="htask_task_001",

  evidence=[

    {"type": "test", "description": "15 unit tests passing"},

    {"type": "review", "description": "PR #123 approved"}

],

  result={"files_modified": ["src/auth/jwt.ts"]}

Verify Before Closing Parent

rlm_htask_verify_closure(swarm_id="swarm_abc", task_id="htask_ws_001")

# Returns:

{"can_close": false, "blockers": ["2 children not completed"], "needs_waiver": true}

Close with Waiver

rlm_htask_close(

  swarm_id="swarm_abc",

  task_id="htask_ws_001",

  waiver_reason="Error handling moved to next sprint",

  waiver_approved_by="tech-lead"

Observability: Metrics & Audit Trails

Real-Time Metrics

rlm_htask_metrics(swarm_id="swarm_abc", period_hours=24)

# Returns:

  "total": 45,

  "by_status": {"pending": 10, "in_progress": 8, "blocked": 3, "completed": 24},

  "throughput": {"closures_total": 15, "per_hour": 0.63},

  "aging_days": {"N0": 14.5, "N1": 7.2, "N2": 3.1, "N3": 1.8},

  "blocked_recovered_ratio": 0.85

Audit Trail

Every action is logged for compliance and debugging:

rlm_htask_audit_trail(swarm_id="swarm_abc", task_id="htask_001")

# Events: create, update, block, unblock, complete, close, delete, waiver

Checkpoint Deltas

For dashboards and incremental sync:

rlm_htask_checkpoint_delta(swarm_id="swarm_abc", since="2026-03-19T09:00:00Z")

# Returns events, closures, blocks since checkpoint

Complete Tool Reference

Tool	Purpose
`rlm_htask_create`	Create task at any level (N0-N3)
`rlm_htask_create_feature`	Create N1 with standard workstreams
`rlm_htask_get`	Get task with children
`rlm_htask_tree`	Get full hierarchy tree
`rlm_htask_update`	Update task fields (whitelist by status)
`rlm_htask_block`	Block task with typed blocker payload
`rlm_htask_unblock`	Unblock and re-evaluate ancestors
`rlm_htask_complete`	Complete N3 with evidence
`rlm_htask_verify_closure`	Check if parent can close
`rlm_htask_close`	Close parent (with optional waiver)
`rlm_htask_delete`	Soft/hard delete task
`rlm_htask_recommend_batch`	Get prioritized N3 tasks ready to work
`rlm_htask_policy_get`	Get swarm policy configuration
`rlm_htask_policy_update`	Update policy (admin only)
`rlm_htask_metrics`	Get throughput, aging, recovery metrics
`rlm_htask_audit_trail`	Get task event history
`rlm_htask_checkpoint_delta`	Incremental sync since timestamp

Best Practices

Use consistent owners

Assign clear ownership at each level for accountability

Define acceptance criteria upfront

Agents need clear "done" definitions to complete correctly

Require evidence for N3 tasks

Prevents premature closure and ensures quality

Mark non-critical tasks as non-blocking

Use is_blocking=false for nice-to-haves

Document all waivers

Future audits will thank you

Monitor blocked_recovered_ratio

Low ratio indicates systemic blockers needing attention

Read Full Documentation Try It Now