Agents Architecture Learning Systems Security Stack Contact
Agentic AI Engineering Autonomous Systems Production Infrastructure

Agents that operate,
not just respond

We build autonomous agent systems that run your operations. Deep workflow integration, persistent memory, multi-model orchestration, and continuous self-improvement. Production infrastructure, not prototypes.

Discuss your project
How our agent systems work
A network of specialized agents with defined responsibilities. Agents communicate bidirectionally — delegating, reporting, requesting re-evaluation. Every agent reasons independently within its domain.
Orchestrator
Routes tasks between agents. Decides execution order, handles retries, and manages inter-agent dependencies.
ORCH
Orchestrator
routes & decides
Ingest Agent
Monitors email, APIs, and file systems. Extracts structured data from unstructured inputs for downstream agents.
IN
Ingest
email, api, files
Analyst Agent
Interprets ingested data against vault context. Identifies priorities, extracts requirements, and generates structured briefs.
ANL
Analyst
context & reason
Memory Agent
Reads and writes persistent vault files. Maintains relationship history, decision logs, and operational state across sessions.
MEM
Memory
vault read/write
Research Agent
Performs deep analysis passes on vault data. Surfaces non-obvious patterns, cross-references external sources, generates novel insights.
RES
Research
deep analysis
Executor Agent
Produces deliverables within spec — drafts documents, generates content, populates templates, executes multi-step actions.
EXE
Executor
draft & generate
Auditor Agent
Evaluates outputs against quality criteria. Has absolute veto power — failed outputs trigger automatic regeneration with feedback.
QA
Auditor
veto & regen
Delivery Agent
Formats final outputs, pushes to target systems, sends notifications through configured channels (email, Telegram, Slack).
OUT
Deliver
deploy & notify
Scheduler Agent
Triggers operations on defined cadences. Compiles briefings, monitors conditions, and initiates proactive workflows without human prompts.
CRON
Scheduler
proactive ops
orchestration
task complete
memory sync
QA reject → regen
Modular agent topology
Scales to any operation size
Bidirectional communication
Agents delegate, report, and re-evaluate
Zero human prompts
Fully autonomous after deployment
Chatbot vs. agent system
Most people have used ChatGPT. They've had mixed results. An agentic system is fundamentally different — it doesn't wait to be asked.
Generic LLM Chatbot
Can you update the client report with last week's numbers?
I don't have access to your files or data. You could copy-paste the numbers here and I can help format them.
What about the feedback from the last review?
I don't have any context from previous conversations. Could you share the feedback again?
Can you send the updated version to the team?
I'm not able to send emails or access your tools. I can draft a message for you to copy.
Waits. Forgets. Can't act.
Integrated Agent System
5 agents active
Memory
Persistent vault. Remembers your client's preferences from six weeks ago.
Workflow
Multi-step pipelines. Ingests, analyzes, executes, inspects, delivers — end to end.
Integration
Direct API connections to your CRM, email, calendar, documents. Acts inside your tools.
Quality
Auditor agents with veto power. Failed outputs rejected and regenerated automatically.
Operation
Runs on a schedule. Scans, surfaces insights, sends briefings. Reaches out to you first.
Remembers. Connects. Acts autonomously.
Core patterns we deploy
Composable building blocks, not rigid templates. Every deployment is composed differently based on the operation.
01

Multi-Agent Pipelines

Specialized agents with enforced handoffs, quality gates, and automatic regeneration on failure.

02

Proactive Cron Operations

Scheduled agents that initiate actions on cadence. Context logs bridge autonomous and interactive sessions.

03

Vault-as-Memory

Persistent knowledge in structured markdown. Human-auditable, version-controlled, zero migration.

04

MCP Integration Layer

Unified protocol for agent-to-service communication. Swap providers without system changes.

pipeline.yaml
# multi-agent pipeline with quality gates # each agent owns one responsibility pipeline: document_generation trigger: "on_demand | scheduled | webhook" agents: - analyst: model: claude-opus-4 role: "Analyze input, extract requirements, assess context from vault memory" output: structured_brief.json - creative_director: model: gemini-2.5-pro role: "Select strategy, define constraints, choose techniques from rules DB" reads: vault/intelligence/rules_db.md - generator: model: gpt-4o role: "Produce deliverable within spec" max_retries: 3 - inspector: model: claude-opus-4 role: "Evaluate against quality criteria. Absolute veto. Rejection → regen." on_fail: regenerate(generator, feedback=inspector.report) - deployer: model: claude-haiku-4 role: "Format, optimize, deliver" quality_gate: min_score: 0.85 auto_improve: "failures → rules_db"
# proactive cron operations # agents that initiate, not just respond schedule: morning_briefing: cron: "0 7 * * *" agent: briefing_compiler actions: - scan pipeline for stale items - check contacts for overdue follow-ups - pull data from integrated APIs - cross-ref calendar with history - compile prioritized briefing - deliver via Telegram / email writes: vault/ops/context_log.md idempotency: stamp-based contract_monitor: cron: "0 9 * * 1" agent: contract_tracker actions: - scan active_terms.md for expirations - flag contracts within 30-day window - check clause conditions against vault - generate renewal recommendations on_finding: escalate_to_orchestrator insight_engine: cron: "0 20 * * *" agent: deep_research actions: - cross-ref all vault domains - surface timing correlations - detect workflow bottlenecks - write findings to novel_insights.md idempotency: stamp-based health_check: cron: "*/30 * * * *" agent: system_monitor alerts: telegram + email
# vault-as-memory architecture # structured markdown = persistent memory vault/ contacts/ client_jordan_kim.md # history client_sarah_chen.md # email sync vendor_atlas_corp.md # terms lead_prospects.md # pipeline operations/ pipeline.md # stages context_log.md # bridge decision_log.md # traces daily_briefings.md # cron intelligence/ follow_up_queue.md # scored novel_insights.md # signals rules_db.md # auto-learn failure_log.md # QA rejects contracts/ template_service.md active_terms.md # tracking renewal_queue.md # alerts config/ agent_rules.md # bounds security_policy.md # access model_routing.md # models # git-versioned. human-auditable. # every write is a commit. # roll back any decision instantly.
# MCP integration layer # agent ↔ external services mcp_servers: google_workspace: protocol: mcp-stdio capabilities: - gmail.read - gmail.draft - gmail.send - calendar.read - calendar.create - drive.search - drive.read document_generation: protocol: mcp-stdio capabilities: - contracts.generate - contracts.track - reports.compile - cma.generate - templates.populate communication: protocol: mcp-stdio capabilities: - telegram.send - telegram.receive - slack.post - sms.send crm_bridge: protocol: mcp-http capabilities: - contacts.sync - deals.update - pipeline.read # standard protocol. swap providers freely. # add new integrations without code changes.
Karpathy auto-learning
Self-improving agent systems that compound intelligence every cycle. The intelligence is in the memory, not the model weights.

Every failure makes the system permanently smarter

Inspired by Karpathy's vision of autonomous self-improvement. Our agents close their own feedback loops — failures update rules, successes reinforce patterns, deep research surfaces what humans miss. No retraining. No fine-tuning.

1
Agent executes task using vault context + rules DBFull context from persistent memory
2
Auditor evaluates against quality criteriaAbsolute veto power on all outputs
!
Failure → regen with feedback injectionAuditor report drives improvement
3
Failure class extracted → new rule writtenrules_db.md grows from real ops data
4
Nightly deep research scans vaultCross-domain patterns, timing correlations
Next run is smarter — compounding, no retrainingSame mistake never happens twice
Rules DB
Self-Evolving
+rules every cycle
Execute
🔍
Audit
Learn
🔬
Research
Update
How we build agent systems
I

Workflow Audit

Map every tool, channel, and decision point. Interview the people doing the work. Understand the domain before writing code.

II

Agent Architecture

Design agent topology, vault schemas, quality gates, model selection, integrations, and human-controlled boundaries.

III

Deploy & Harden

Ship to production. VPS-hosted, always-on. Interface through existing tools. Continuous self-improvement from day one.

Production-grade agent systems
Same composable architecture, tailored to different domains. We adapt these patterns to any profession with complex workflows.
01

Enterprise Workflow Automation

Vault + Cron + MCP + Pipeline

Organizations running on fragmented tools with no central intelligence layer. Scheduling, finance, communications, and project management scattered across disconnected systems.

Multi-channel ingest
Cross-domain orchestration
Signal-based scheduling
Proactive briefings
12+scheduled cron ops
100+persistent vault files
9MCP integrations
24/7autonomous
Architecture: Full-stack vault-as-memory with Obsidian-compatible knowledge base. Multi-channel ingest scans email, calendar, messaging, and document repos. Orchestrator coordinates cross-domain actions — scheduling changes affecting resource allocation, financial events triggering reporting, stakeholder updates cascading through channels. Signal-based adaptation over fixed schedules. Stamp-based idempotency across cron cycles. Proactive outreach as primary interaction mode.
02

Autonomous Content & Document Pipeline

Multi-Agent Pipeline

Professional content requiring deep domain knowledge — contracts, market analyses, marketing materials, technical documentation. Generic AI lacks the domain depth to be production-ready.

6-agent pipeline
QA with absolute veto
Self-improving rules
Multi-format output
6pipeline agents
7quality criteria
19technique library
~minvs days production
Architecture: Six-agent pipeline with enforced handoffs. Analyst extracts requirements and domain context from vault. Director selects strategy from evolving rules database. Generator produces deliverable — contracts, CMAs, reports, marketing assets — using domain-specific models. Auditor evaluates against 7 criteria with absolute veto. Failed outputs auto-regenerate with feedback injection. Self-improving rules database: every QA failure updates constraints for future runs.
03

Operations Management Agent

Vault + Cron + MCP

Professional services operations spanning dozens of active engagements. Communications, contracts, expenses, scheduling, follow-ups, and reporting all running through one operator's memory.

CRM integration
Contract generation
Expense tracking
Follow-up intelligence
12daily operations
25+agent actions
3,600+lines orchestration
8services
Architecture: Vault-as-memory with per-client files auto-synced from email, calendar, and voice transcriptions. CRM integration for contact lifecycle, deal progression, and status updates. Contract engine populates templates, tracks term expiration, flags renewal windows, monitors clause conditions. Expense management with receipt processing. Follow-up intelligence scores urgency from relationship history and time decay. Morning briefings, evening recaps, deep research surfacing cross-client patterns.
Built for organizations that need
to trust their AI
Hard boundaries, full auditability, zero cloud dependency on sensitive data.

Local-First Architecture

All vault data and logs on your infrastructure — VPS or on-premise. No sensitive data transits third-party storage. You own the data.

Multi-Model Orchestration

We use the latest frontier models and optimize each agent role for reasoning depth, execution speed, and cost. No single-provider dependency — swap models without system changes.

Hard-Coded Security Rules

Behavioral constraints in code, not prompts. Data boundaries, action limits, and escalation triggers enforced at the system level.

Full Audit Trail

Every decision logged with reasoning traces. Git-versioned memory — see what changed, when, and why. Roll back any action instantly.

NemoClaw-Class Guardrails

Policy-based privacy and security controls inspired by NVIDIA NemoClaw. Agent behavior boundaries, data handling rules, and controlled cloud access with local-first processing for sensitive operations.

Privacy Router

Sensitive operations run locally. Cloud-routed tasks filtered through defined security boundaries. Per-agent data access scopes with no lateral privilege escalation.

Technology
Best tool for each layer. Swap components without system rewrites.
Claude Opus / Sonnetreasoning
GPT-4omulti-modal
Gemini 2.5reasoning
Claude Code CLIagent runtime
MCPintegration
Pythonorchestration
TypeScriptpipelines
Bashautomation
NVIDIA NemoClawguardrails
Hetzner VPSinfra
Cloudflaredns
Google Workspaceemail & docs
DocuSigncontracts
Telegram / Slackagent interface
Obsidianvault ui
Gitversioning
Veo / Klingvideo gen
ElevenLabsvoice
Remotionmedia
Puppeteerbrowser
FFmpegprocessing
Cron / Systemdscheduling

Vecreal is a Canadian AI engineering firm based in Alberta. We build autonomous agent systems for organizations that need AI to do real work — operate pipelines, manage relationships, produce deliverables, generate contracts, and make decisions within defined boundaries.

Our systems are in production. Every pattern we deploy comes from real deployments, not theory. We work with a small number of clients at a time to ensure depth.

For inquiries about agent system development:
admin@vecreal.com