From cornmill.online
Cornmill Agentics builds autonomous AI agents modelled on biological cognitive systems. Our agents don't just respond — they learn, remember, sleep, and evolve.
Design Philosophy
3.8 billion years of evolution have produced the most sophisticated information processing systems known to exist. Every biological organism is a solution — refined across countless generations — to the problems of perception, memory, decision-making, and adaptation.
Modern AI largely ignores this inheritance. Most systems are stateless — brilliant in the moment, amnesiac by design. They never sleep, never consolidate, never develop the institutional memory that makes biological intelligence so powerful.
We take a different approach. Rather than engineering from first principles alone, we study the patterns that evolution has already validated and translate them into agent architectures. Sleep cycles that consolidate memory. Specialisation that mirrors ecological niches. Trust hierarchies borrowed from social species. Delegation patterns found in colony organisms.
The result: AI agents that don't just process — they grow.
Why solve from scratch what nature perfected over eons? We extract proven cognitive patterns from biology and implement them in silicon — accelerating development by leveraging billions of years of R&D.
Biological brains don't just store — they prune, merge, and synthesise during sleep. Our agents do the same, transforming raw interactions into distilled knowledge through dream cycles.
In nature, species find niches. In our system, agents develop distinct expertise and personalities, delegating to specialists just as organisms in an ecosystem share the cognitive load.
Bio-Mimetic Design
During sleep, the brain cycles through light, deep, and REM stages — pruning weak connections, strengthening important ones, and discovering cross-domain patterns.
Agents accumulate "melatonin" through activity. When thresholds are reached, they enter dream cycles — light sleep prunes redundancy, deep sleep consolidates, REM discovers cross-user insights.
Organisms regulate activity cycles through chemical signals — melatonin rises with sustained wakefulness, triggering the need for rest and consolidation.
Digital melatonin accumulates with each message, tool use, and memory operation. The system self-regulates — high engagement naturally triggers consolidation periods.
Ant colonies, bee hives, and neural networks achieve complex behaviour through specialised agents communicating via simple protocols — no central controller required.
Specialised agents delegate tasks to each other, preserving context. Like neurons in a brain, each handles its domain while the collective achieves far more than any individual.
Social species develop trust through repeated interaction — from stranger to ally. Access to shared resources scales with established trust.
Every communication channel carries trust tiers. New contacts start restricted; verified senders gain tool access. The agent's capabilities scale with the relationship.
The System
cornOS is our core platform — a cognitive agent orchestration system that deploys AI assistants capable of genuine learning and autonomous operation. Explore the full platform →
Agents monitor multiple channels — email, messenger, webhooks, scheduled triggers — ingesting messages with full context and sender trust verification.
Every interaction is informed by persistent memory — per-user preferences, global knowledge, and document intelligence via RAG. No conversation starts from zero.
Agents plan and execute multi-step operations using 26+ tools — sending emails, searching documents, delegating to specialists, managing calendars, and more.
After sustained activity, agents enter bio-mimetic dream cycles. They prune, merge, and synthesise — waking up sharper, with distilled knowledge and cross-domain insights.
Capabilities
Per-user and global memory that persists across conversations. Agents remember preferences, context, and learned patterns — building genuine relationships over time.
Bio-mimetic sleep stages — light, deep, REM, and lucid — that consolidate experience into knowledge. Agents don't just store data; they learn from it.
Specialised agents hand off tasks to each other with full context preservation. Local and remote delegation via the Model Context Protocol.
Unified agent identity across email, messenger, WhatsApp, SMS, Telegram, and webhooks — with trust-aware permissions per channel.
Per-agent RAG system with semantic chunking and embedding search. Agents can ingest PDFs, manuals, and policies, then reference them contextually.
From email and calendar to file management, web search, OCR, AppleScript automation, and financial operations — plus unlimited extensibility via MCP.
Cron-based task scheduling lets agents operate independently — running daily summaries, periodic maintenance, and custom automated workflows.
Built-in approval gates for sensitive operations. Trust policies ensure agents act within defined boundaries while maintaining autonomous capability.
Research Applications
Run concurrent agent populations with controlled differences to model outcomes — turning theory into observable, repeatable experiment.
Because our agents develop genuine behavioural characteristics — memory, personality, trust dynamics, learning rates — they become something unprecedented: controllable proxies for studying complex adaptive systems.
Deploy two identical agent populations, vary a single parameter — management style, information flow, incentive structure — and observe how outcomes diverge over time. The bio-mimetic architecture means these aren't abstract simulations; the agents learn, adapt, and exhibit emergent behaviours just as biological systems do.
Model the effects of flat vs hierarchical structures, different delegation strategies, or varying autonomy levels. Observe how information flows, decisions propagate, and institutional knowledge develops under each configuration.
Simulate market participants with persistent memory and learning capability. Study how trust dynamics, information asymmetry, and incentive design affect emergent market behaviour — with agents that genuinely adapt rather than follow scripted rules.
Test hypotheses about cooperation, specialisation, and adaptation. Because the agents' cognitive architecture mirrors biological systems, observed dynamics map meaningfully to evolutionary theory — at a pace biology can't match.
Study how information spreads, distorts, and consolidates across agent networks. Model the effects of different communication topologies, trust thresholds, and channel configurations on collective intelligence.
Define your agent populations with identical base parameters. Identify the variable you want to study.
Introduce a single controlled difference — a management structure, a memory parameter, a trust threshold, an incentive rule.
Run both populations concurrently through identical scenarios. The system logs every decision, memory operation, and interaction.
Compare outcomes across populations. Because agents are bio-mimetic, divergences in behaviour map to meaningful real-world hypotheses.
Case Study
Proof that autonomous agents can run real-world products — not just answer questions.
Every morning, with no human intervention, our agent system researches, writes, edits, and publishes a complete daily newspaper — now on its 41st edition and counting.
The Cornmill Intelligencer is a fully autonomous publication covering Scottish and UK news, science, community affairs, wildlife, audio technology, and more. It features a front page, dedicated sections, a "Good News Index" tracking positive stories, and even a playful Page 3 featuring bird photography.
This isn't a demo or a proof of concept. It's a live product, updated daily, read by real people. It demonstrates what becomes possible when AI agents have persistent memory, scheduled autonomy, and the ability to coordinate complex multi-step workflows without human oversight.
Read Today's EditionWhether you're exploring autonomous agent systems for your organisation or interested in the research behind our approach, we'd love to hear from you.