The model layer will keep changing.
Your operations platform shouldn't.
Seven structural shifts are coming to the AI industry. We've built Opsilon to perform in all of them, because we made different architectural choices from day one.
Most AI tools are built for the model landscape of today. They assume a specific cost structure, a specific capability tier, a specific data residency posture. When the model layer shifts, and it will, repeatedly, those tools have to rebuild, or you have to migrate.
Opsilon is built around what won't change: the need to govern agents, to own institutional memory, to coordinate humans and AI as peers, and to answer the compliance question your auditors will ask. Seven predictions. Seven architectural decisions we made before they were obvious.
AIs will get smarter.
Frontier model capability has compounded consistently. Benchmarks that required careful scaffolding in 2023 are now trivial. Every few months a new model raises the floor for what agents can do autonomously.
As models get more capable, agents become more autonomous. More autonomy means higher stakes on every decision, and a greater need for the infrastructure to audit, gate, and govern what agents do. The question stops being 'can AI do this?' and becomes 'do I have the infrastructure to trust AI doing this at scale?'
Model-agnostic architecture means every capability improvement is a free upgrade, route any workspace to any model, today or in two years. The governance layer (reasoning log, audit trail, human gates, agent versioning) becomes more valuable as the model gets more capable. The memory and SOPs your agents have accumulated compound. The harness layer gets simpler as models improve, exactly as the research shows.
We build the layer above the model, not the model itself. Every improvement to the model layer improves Opsilon automatically.
AIs will get cheaper per token.
Inference costs have dropped orders of magnitude since GPT-3. Competition, efficiency research, and commoditisation of model serving all push in the same direction. This trajectory continues.
Tasks that weren't worth automating at $10/run become viable at $0.10/run. The bottleneck shifts from 'can AI do this?' to 'can I deploy this safely across the operation?' Budget governance, per-agent, per-task spend limits, becomes the CFO's primary control surface, not just an API rate limit a developer manages.
Per-agent and per-task budget controls are structural, not an afterthought. Route each role to the cheapest model that meets the quality threshold, fast cheap model for triage, more capable model for reasoning, local model for sensitive data. As inference costs fall, the same governance infrastructure governs an expanding set of tasks without a rebuild at each cost threshold.
Cheap inference + governed deployment = the entire operation becomes automatable. Not just the pilot workflow.
AIs will get more efficient in token use.
Models are rapidly improving at doing more with less, better instruction-following, compressed reasoning, more reliable tool use in fewer turns. Context window costs are falling alongside capability improvements.
Agents that required long, heavily scaffolded prompts will run leaner. Harnesses that pre-loaded every skill into every session will look wasteful. Token efficiency compounds, a more efficient model in a more efficient harness is a step-change in operating cost and response latency.
Skills are lazy-loaded per task context on demand, not pre-loaded into every session. Task-isolated contexts mean each worker starts clean with no context pollution from unrelated prior work. Memory is structured and queryable, not a flat file the agent reads in full. As models get more efficient, Opsilon's architecture gets cheaper to run, not more expensive.
We designed for efficiency from the start, not as a cost-reduction project after the fact.
AIs will go local.
Regulatory pressure on data residency is intensifying, EU AI Act, ASIC, sector-specific financial services rules. The model capability required to run useful agents on local hardware is arriving. Several open-weight models already meet the quality threshold for back-office operations tasks.
For PE funds, law firms, and accounting practices, 'our deal data transits OpenAI's infrastructure' will become an increasingly untenable answer to a compliance question. Local inference, on-device, in a private VPC, or via a self-hosted open-weight model, will be a procurement requirement, not an edge case.
Model-agnostic routing supports any inference endpoint, Anthropic, OpenAI, Azure, AWS Bedrock, or your own in-VPC model today. Self-hosted deployment is on the roadmap. When local models meet the quality threshold for a given task type, you route there. The governance layer, memory substrate, and audit trail are unchanged. No migration, no rebuild.
The model runs where the data policy requires. The platform doesn't move.
Data control becomes the #1 priority for every business.
LangChain's own analysis named memory lock-in as the primary moat in the agent category. Anthropic Memory, ChatGPT Workspace Agents, and every platform building persistent agent memory knows this. Whoever holds your agents' accumulated knowledge holds your operation.
The institutional knowledge encoded in your agents, SOPs, counterparty histories, screening criteria, learned corrections, is the compounding asset. If it lives in a vendor's runtime, switching means starting over. Regulatory pressure compounds this: GDPR, sector data localisation requirements, and ISO 42001 all push toward owning your own AI data substrate.
All memory lives in your Postgres, not Anthropic's runtime, not OpenAI's servers, not a proprietary vendor store. Customer-managed keys (BYOK) and hold-your-own-key (HYOK) on Enterprise. AU data residency by default. No cross-tenant memory. Zero-data-retention agreements apply with all model providers by default. Your agents' accumulated knowledge is yours to keep, migrate, audit, and export.
Memory is the moat. It should be your moat.
Domain-tuned models will outperform generalists on professional work.
Harvey, MikeOS, and Bloomberg's BloombergGPT aren't experiments, they're early evidence of a structural pattern. Models trained on dense domain corpora (case law, deal documents, financial filings) consistently outperform frontier generalists on vertical tasks, often at lower cost. The moat isn't scale anymore; it's specificity.
The model choice for a given professional task won't be 'which frontier model?', it will be 'which frontier model, and which domain-tuned model for this specific context?' Legal work will route through law-trained models. Financial analysis through finance-trained models. The correct answer will differ by task type, client context, and regulatory jurisdiction. Locking your operation into one model provider, or one model type, is an increasingly bad bet.
Opsilon's model-agnostic routing layer was built for exactly this: as domain-tuned models reach quality thresholds for specific task types, you route there without changing the governance layer, the memory substrate, or the audit trail. The same approval chains, reasoning logs, and agent versioning work regardless of what's under the hood. Your operation doesn't have to pick a side in the model wars, it benefits from all of them.
The platform is the constant. The models it uses will keep improving and specialising.
Vertical agent SaaS turns your operations into a commodity.
When every firm in your sector uses Harvey, and Harvey is trained on the same legal corpus, the output converges. Speed improves across the industry simultaneously. The tool becomes a table stake, not an edge. This is already visible: early Harvey adopters no longer talk about it as a competitive advantage, they talk about it the same way they talk about their document management system.
The two assets that actually differentiate professional services firms are proprietary data, deal records, client histories, case outcomes, institutional memory, and proprietary processes, how you screen, how you structure, how you decide. Both live outside vertical SaaS. They exist in your people's heads, your email threads, and your tribal knowledge. If they don't get encoded anywhere, they walk out the door with your senior staff.
Opsilon is where your processes get encoded and your institutional memory accumulates. Your screening criteria become agent instructions. Your approval chain becomes a structural gate. Every task completed adds to a memory substrate in your infrastructure, deal context, client notes, learned corrections. A firm running Opsilon for two years has agents that know exactly how they work. A firm running Harvey for two years has faster analysts doing the same work as everyone else.
The model is a commodity. Your process and your data aren't. Encode them somewhere you own.
Whatever shifts, you're covered.
Seven predictions. Seven architectural decisions we made before they were obvious. Model-agnostic from day one. Memory in your infrastructure. Governance-first. Opsilon isn't a bet on one model, one cost structure, or one regulatory environment.
See how it works →