CAPABILITY
AI Capability
Models are good enough. The harness isn't.
A structured view of a fast-moving field, written for solo operators, small practices, and the AI-curious in micro-organisations using AI as part of how the work gets done.
The model layer commoditises every six months. Tools (harnesses) converge. What compounds is the structure you build around the model: your context, your skills, your saved instructions, your library of work. We call that the AgentOS, and it's what we help you build.
MAY 2026 LENS
Four pressures now shape practical AI capability
The framework still holds, but the field has tightened around four live pressures: scarce tokens, staged agent work, richer interaction, and human value that remains part of the service.
Scarcity
Compute, tokens, attention, and human review are now planning constraints, not background details.
Staging
More work now happens before execution: setting context, constraints, rubrics, handoffs, and success criteria.
Interaction
Voice, pointing, interruption, browser context, and real-time feedback are becoming model capabilities.
Human Premium
Trust, accountability, presence, translation, behaviour change, and provenance still carry economic value.
Deployment capacity cuts across all four tiers. Compute and infrastructure sit in the Landscape. Pricing, usage limits, governance, and support capacity sit in the Foundation. Staging, review loops, and routing habits sit in the Practice. Service redesign, human escalation, and continuous support sit in the Application. It is not a fifth tier; it is the constraint that now runs through the system.
An AI Framework
We organise AI capability into four tiers: the landscape outside, the foundations that hold it up, the practice where work happens, and the application that delivers outcomes. Each tier maps to specific layers of the AgentOS framework. The tiers are the public map; AgentOS is the operating architecture underneath.
The Landscape
Model and provider moves, compute capacity, interaction models, infrastructure, and the tools now available.
Models, Harnesses, Tools & Platforms
The landscape tracks what is available and what is changing: frontier models, work harnesses, coding and design environments, voice and interaction layers, automation platforms, integration standards, and the compute infrastructure that now shapes limits, latency, pricing, and reliability.
Explore the landscape →The Foundation
The static AgentOS layers — who you are, what's around you, what you connect to
Identity, Context & Connections (plus governance, readiness, trust)
Three of the seven AgentOS layers live here: Identity (who you are, what you do, what you stand for), Context (your situation, what's true now — Pandion calls this layer's discipline Context Engineering), Connections (data sources, services, tools your AI can reach — with trust grades and access controls). Strategy & governance set direction. Token economics, usage limits, deployment bottlenecks, privacy, sustainability, and risk controls decide whether the system can be used safely and repeatedly.
Explore the foundation →The Practice
Staging work, handoffs, rubrics, agent management, verification, and interaction as context transfer.
Skills, Memory, Verification & Team Fluency
Three more AgentOS layers in operation. Skills — procedural knowledge made portable; agents that can be loaded and run. Memory — what compounds across sessions; what gets remembered, summarised, archived. Verification — trust, evaluation, grounding so outputs hold up. The May shift is from prompting to staging: context, constraints, rubrics, mixed-doneness handoffs, review loops, and routing habits that let humans manage agents rather than just prompt them.
Explore the practice →The Application
Human premium, service-model change, new roles, continuous support, and outcomes in a real domain.
Automations & Outcomes in your Domain
The seventh AgentOS layer (Automations) plus the point of all of it (Outcomes). AI applied to specific domains: sustainability reporting, business systems, research synthesis, knowledge management, strategy and planning. This tier asks what becomes cheaper, more continuous, more personalised, or newly possible, and where trust, accountability, relationship, and human escalation still matter.
Explore applications →ANOTHER LENS
Same picture, layered
The four tiers above describe where the work happens. Stepping back, here's the same picture seen as a layered system: what runs, what reads what, and what compounds across tools. The framework is how to orient; AgentOS is how to operate.
Models commoditise. Harnesses converge. The AgentOS — seven persistent layers written as plain text files at the root of your workspace — is the part that compounds across both, survives every tool swap, and is genuinely yours. This is what we build.
Layer 1
MODEL
Commoditising
The reasoning engine. GPT, Claude, Gemini, Llama. Pretrained knowledge + language understanding. Picks change quarterly; you don't marry one. (Sits in Tier 1, the Landscape.)
Layer 2
HARNESS
Converging
The runtime engine. Claude Code, Cursor, Codex. Software that runs the model, calls tools, manages state. You probably already use one. They're becoming interchangeable. (Also in Tier 1, the Landscape.)
Layer 3
AGENTOS
Compounding
The persistent foundation. Plain text files at the root of your workspace, read by every harness at every session start. Seven layers, each a discipline:
1. Identity
Who you are. What you do. What you stand for. The file every other layer references. (Tier 2, Foundation.)
2. Context
Your situation, what's true now, what your operating environment looks like. Pandion calls this layer's discipline Context Engineering. (Tier 2, Foundation.)
3. Skills
Procedural knowledge made portable. Agents and capabilities that can be loaded and run. (Tier 3, Practice.)
4. Memory
What compounds across sessions. What gets remembered, summarised, archived. (Tier 3, Practice.)
5. Connections
The data sources, services, and tools your AI can reach. Trust-graded, access-controlled. (Tier 2, Foundation.)
6. Verification
How outputs are checked, grounded, evaluated. Trust by construction, not by hope. (Tier 3, Practice.)
7. Automations
What runs without you. Scheduled jobs, triggers, agents that act on signal. Where the AgentOS does work in the background. (Tier 4, Application.)
These seven layers are what we build. Plain text. Yours. Portable across every tool.
Layer 4
OUTCOMES
The point of it all
Real-world impact. Better decisions, faster cycles, capability that grows session over session, your team operating differently. (Tier 4, Application.)
Example in practice
What this looks like in a Pandion engagement
For embedded engagements with solo operators and micro-orgs, we deploy the full infrastructure stack:
- • Multi-party AgentOS architecture — per-party engagement folders that scale as your team grows (you, future hires, contractors, vendors)
- • Embedded skills — an orchestration skill for your daily work, plus a Pandion-side skill that stewards the system and stands in when we're offline
- • Chief-of-staff weekly review — synthesis across sessions, surfaces what shipped and what is open
- • Block Hours commercial model — no monthly expiry, transparent ledger, top-up when balance runs low
Reusable methodology. Ships with every embedded engagement.
Explore the Framework
See how tools, concepts, and domains connect across the four tiers. Click any node to dive deeper.
Loading knowledge graph...
In 30 Seconds
The gap between people using AI effectively and people still experimenting is no longer widening, it has become a chasm. Not because the technology is inaccessible (everyone has access now), but because structure compounds. Solos and small practices that moved early on the AgentOS layer are now operating fundamentally differently.
We help you:
- Navigatethe model + harness landscape (what to use, what to skip, what just shipped)
- Buildyour AgentOS — the persistent foundation that compounds across tool changes
- Developpractice — context engineering, skills, the working pattern with AI
- ApplyAI to your domain — visual work, scheduled work, long-document work, and the work specific to your practice
Most disappointing AI output is a context failure, not a capability failure. AI agents can run for hours on a well-briefed task. Most people's actual median use is forty-five seconds. The gap isn't access. It's the structure you've put around the model and how clearly you brief it.
Where to start: a working AgentOS, however small. One identity file. One context file. A short list of standing instructions. The structure pays back across every prompt for the next quarter and travels with you when the tools change.
Start with the Readiness Assessment →How we work: we don't just advise. We run an AgentOS ourselves and have for two years. We build, audit, and own the rhythm so it doesn't go stale. That keeps us current and lets us deliver real value at rates that work for one-person businesses.
The Challenge
For solo operators & founders
Running your own show. Curious about AI, time-poor, building or running a small business. You've tried Claude or ChatGPT, maybe a few tools. Results are real but patchy.
- • Which tools are real vs hype for one person?
- • How do I get consistent value from AI, not one-off tricks?
- • Where do I start with a working AgentOS, however small?
For small domain-led practices
Landscape consultancies, conservation orgs, architects, accountants, working farms, hospitality. Domain expertise first, AI as a tool inside the practice. Wanting AI in your voice, not generic AI voice.
- • How do we make AI write and decide in our voice?
- • How do we keep client-confidential data safe while using AI broadly?
- • What stays bespoke and what becomes templated?
The landscape overwhelm
The options are exploding: Claude, ChatGPT, Gemini, Copilot, Cursor, Codex, ten new no-code platforms. New vocabulary every month: harness, AgentOS, Markets-of-One, monothread. You don't have time to track all of it.
- • What's actually different about Opus 4.7, Claude Design, GPT-5.5?
- • What does the new vocabulary mean for one person?
- • Who can I trust for guidance that isn't selling something?
The gap: most disappointing AI output is a context failure, not a capability failure. The model is good enough. What's missing is the structure around it.
Tools without an AgentOS are toys. AgentOS without rhythm goes stale in eight weeks. And solos most often stall on the structure work, not the technology.
AI Benefits & ROI
AI can deliver value in many ways. Understanding the different types of benefit helps you set realistic expectations and measure what matters.
Time Savings
Faster completion of routine tasks - drafting, research, analysis, admin.
Cost Reduction
Lower cost per output through automation and efficiency gains.
Increased Output
More work completed with the same resources - volume and throughput.
Quality Improvement
Better accuracy, consistency, and fewer errors in deliverables.
Revenue Growth
New revenue streams, better conversion, or enhanced offerings.
New Capabilities
Doing things you couldn't do before - analysis at scale, synthesis, expertise on demand.
Risk Reduction
Better compliance, early warning systems, error catching before problems occur.
Better Decisions
Right information at the right time - surfacing insights that inform action.
ROI isn't just cost savings. Some benefits are easy to measure (hours saved, costs cut). Others - like better decisions or new capabilities - are harder to quantify but often more valuable.
The key is knowing what type of value you're pursuing and measuring accordingly.
How to measure AI ROI →How We Help
Building AI capabilities that actually work - across the full journey
Landscape Navigation
Cut through the noise, find what matters
- •Tool Selection: Which AI tools are right for your context and constraints-
- •Ecosystem Thinking: Partners, vendors, platforms - how they fit together
- •Pace of Change: What's stable, what's shifting, where to bet
More capability ships quarterly than used to ship in years. We help you stay oriented.
Foundation Building
Strategy + governance + data/knowledge readiness
- •AI Strategy: Vision that starts with your reality, not vendor pitches
- •Governance: Risk, ethics, and compliance as enablers, not blockers. Recognises ISO 42001, the EU AI Act, NIST AI RMF, and OWASP LLM Top 10.
- •Privacy by Architecture: Data minimisation, encryption, and access controls that stand up to audit
- •Knowledge Readiness: AI exposes gaps you didn't know you had
Knowledge is more valuable than data alone. Privacy is the trust layer that makes AI usable.
Execution & Fluency
Context, orchestration, capability - our core focus
- •Context Engineering: Right information, right time - fix the context, fix the results
- •Agents & Orchestration: Skills-based systems where expertise compounds - plus operations and governance
- •Team Fluency: Thinking with AI, not just using AI tools
Most AI projects stall here. This is where we focus.
Domain Application
Where AI meets your actual work
- •Sustainability & Reporting: AI-enhanced disclosure, standards navigation, data synthesis
- •Research & Analysis: Landscape mapping, evidence synthesis, competitive intelligence
- •Strategy & Planning: Scenario analysis, theory of change, operating model design
AI without domain expertise produces generic outputs. We bring both.
Who We Help
Solo operators & founders
Running your own show. Curious about AI, time-poor, building or running a small business. Wanting AI in the workflow without becoming an engineer.
How we help:
- • Working AgentOS, however small
- • Repeatable structure that compounds
- • Routing across models without overspending
- • Capability that travels when tools change
Small domain-led practices
Landscape consultancies, conservation orgs, working farms, hospitality, single-handed professional practices (architects, accountants, lawyers, therapists). Domain expertise first, AI as a tool inside the practice.
How we help:
- • AI in your voice, not generic AI voice
- • Public/Private wall for client-confidential data
- • Saved instructions, house style, templates
- • What stays bespoke and what becomes templated
Mission-driven micro-orgs
Small NGOs, social enterprises, charities, conservation organisations. Resource-constrained teams needing to do more with less, without losing the depth that makes the work valuable.
How we help:
- • Research synthesis and document processing
- • Reporting and disclosure assistance
- • Theory-of-change articulation in AI
- • Impact at accessible cost
Personal-life AI users
Often the same person as above in a different hat. Running life admin, property, finance, school and tutoring with AI. Often where the AgentOS habit is most easily learned, then carried into work.
How we help:
- • Personal Context Portfolio
- • Household automations
- • Tutoring and learning support
- • Privacy and data discipline
Navigate + Build
AI CAPABILITY
“How do you work with AI?”
Context, orchestration, fluency - the capabilities
OTHER CAPABILITIES
“Where do you apply it?”
Strategy, systems, capital - AI-enhanced throughout
AI doesn't exist in isolation. It connects to your strategy work, your systems, and your sustainability goals. We work across all of these, because they're connected.
Ready to Talk AI?
Whether you're exploring AI strategy, building AI capabilities, or need specialist expertise to extend your team - let's have a conversation.
No commitment, no pitch. Just a conversation about what might help.
Start with the basics:
- • What are you trying to achieve with AI?
- • What have you tried? What worked or didn't?
- • What would “good” look like?