Real-World Agent Workflow Examples

March 26, 2026 · View on GitHub

Steal these end-to-end flows when building your own automations. Each example shows the goal, prompts, API/CLI calls, and outputs we expect.

1. Feature Development Sprint (BrainMeld PRD excerpt)

Goal: Build "Lessons Learned" field.

  1. Create task 
    vk create "Feature: Lessons Learned field" --project veritas-kanban --type feature --priority medium
  2. Prompt (worker) 
    Implement markdown lessonsLearned field on tasks (UI + API). Include migration + docs. Cross-model review required.
  3. Workflow
    • vk begin <id>
    • Implement server -> shared -> web changes
    • Update docs + tests
    • vk done <id> "Added lessons learned field"
  4. Outputs
    • Task summary with PR link
    • Lessons Learned comment describing future usage

2. Bug Fix (Archive bulk action)

Goal: Sprint archive button fails.

  1. Create bug task referencing GitHub Issue #86.
  2. Subtasks:
    • Reproduce in dev
    • Inspect network requests
    • Patch bulk archive handler
    • Add regression test (Playwright)
  3. CLI flow: vk begin, fix, vk done "Bulk archive now calls API"
  4. Cross-model review ensures UI + API parity.

3. Documentation Update

Goal: Add sanity checks to Getting Started.

  1. Task description includes sections to cover (API, UI, agent pickup).
  2. Agent edits docs/GETTING-STARTED.md + docs/TROUBLESHOOTING.md references.
  3. Completion summary links to diff + screenshot placeholders.

4. Security Audit (RF-002 style)

Goal: Run cross-model audit on repo.

  1. Task -> type=security, project=veritas-kanban.
  2. Subtasks: scope, run Codex audit, run Claude review, compile findings, create issues.
  3. Agents spawn using research prompt template, save results to refactoring/rf-002/*.
  4. Deliverables: Markdown report, HTML deck, GitHub issues.

5. Content Production (Podcast clip → LinkedIn post)

  1. Task type=content with acceptance criteria (summary, caption, schedule time).
  2. Agent fetches transcript, writes summary, drafts LinkedIn copy, saves assets to projects/start-small-think-big/....
  3. Completion summary includes copy + asset path; lessons learned capture platform insights.

6. Research & Report (Champions)

  1. Task type=research, project social, sprint CHAMP-02.
  2. Prompt includes dossier template, required sources, HTML deck requirement.
  3. Agent workflow: gather sources, write Markdown, generate HTML via script, brain-write.sh to mirror.
  4. Final comment: TL;DR + links to both artifacts.

Pattern to Copy

For any workflow:

  1. Task with crystal-clear done definition.
  2. Prompt stored in registry.
  3. API/CLI calls scripted (vk begin/done, time tracking, status updates).
  4. Artifacts saved to predictable paths and mirrored to Brain/engram if needed.
  5. Cross-model review if code/critical.
  6. Lessons learned field updated for systemic knowledge.

Use these recipes as seeds for your own automation playbooks.

7. Workflow Engine Pipeline

Goal: Automate plan → implement → test → review with retry policies.

  1. Create .veritas-kanban/workflows/feature-dev.yml with planner, developer, and tester agents.
  2. Start via API: POST /api/workflows/feature-dev/runs
  3. Monitor live in the Workflows tab — each step shows status, duration, and output preview.
  4. Gate steps block until quality checks pass or a human approves.

See WORKFLOW-GUIDE.md for full YAML examples.

8. Using Task Dependencies

Goal: Ensure backend API is complete before frontend work starts.

  1. Create US-100 "Build REST API" and US-101 "Build React UI".
  2. Set dependency: US-101 depends_on US-100.
  3. The dependency badge on US-101 shows it's blocked until US-100 is done.
  4. Query the full graph: GET /api/tasks/US-101/dependencies

9. Crash-Recovery Checkpointing

Goal: Resume long-running agent work after a crash.

# Save checkpoint mid-work
curl -X POST http://localhost:3001/api/tasks/US-42/checkpoint \
  -H "Content-Type: application/json" \
  -d '{"state":{"step":3,"completed":["auth","db"],"notes":"Working on API layer"}}'

# After restart, resume from checkpoint
CHECKPOINT=$(curl -s http://localhost:3001/api/tasks/US-42/checkpoint)
# Feed $CHECKPOINT into agent prompt for continuity

# Clean up after completion
curl -X DELETE http://localhost:3001/api/tasks/US-42/checkpoint

10. Observational Memory for Cross-Agent Learning

Goal: Capture architectural decisions so future agents don't repeat exploration.

# Log a decision
curl -X POST http://localhost:3001/api/observations \
  -H "Content-Type: application/json" \
  -d '{"taskId":"US-42","type":"decision","content":"Chose WebSocket over SSE for real-time updates — lower latency, bidirectional","importance":9}'

# Future agent searches before making the same decision
curl "http://localhost:3001/api/observations/search?query=websocket+vs+sse"

11. Agent Policy Evaluation (v4.0)

Goal: Restrict an agent from deleting production tasks.

# Create a deny-first policy
curl -X POST http://localhost:3001/api/policies \
  -H "Content-Type: application/json" \
  -d '{
    "name": "no-delete-production",
    "description": "Prevent agents from deleting tasks in production projects",
    "scope": {"project": "production"},
    "rules": [{"tool": "task.delete", "action": "deny", "reason": "Production tasks require human approval for deletion"}],
    "precedence": "deny-first"
  }'

# Test before deploying
curl -X POST http://localhost:3001/api/policies/POLICY_ID/evaluate \
  -H "Content-Type: application/json" \
  -d '{"agent": "codex-1", "tool": "task.delete", "context": {"project": "production"}}'

12. Behavioral Drift Monitoring (v4.0)

Goal: Detect when an agent's task completion rate drops.

# Configure a drift monitor
curl -X POST http://localhost:3001/api/drift \
  -H "Content-Type: application/json" \
  -d '{
    "agent": "TARS",
    "metric": "completion_rate",
    "baseline": 0.85,
    "warningThreshold": 0.70,
    "alertThreshold": 0.50
  }'

# Check drift status across all agents
curl -s http://localhost:3001/api/drift | jq '.data[] | {agent, metric, status}'

13. Decision Audit Trail (v4.0)

Goal: Log a significant architectural decision with assumptions for future reference.

# Record a decision
curl -X POST http://localhost:3001/api/decisions \
  -H "Content-Type: application/json" \
  -d '{
    "taskId": "US-200",
    "agent": "VERITAS",
    "decision": "Use file-based storage instead of SQLite for v4.0",
    "confidence": 0.8,
    "evidence": ["Current scale is <1000 tasks", "File ops are simpler to debug", "No migration path needed"],
    "assumptions": ["Scale stays under 10k tasks", "Single-instance deployment"]
  }'

# Later: record what happened
curl -X POST http://localhost:3001/api/decisions/DECISION_ID/outcome \
  -H "Content-Type: application/json" \
  -d '{"outcome": "File storage held up well through v4.0 launch. Assumption about scale still valid at ~350 tasks."}'