Custom agent systems, run by us
Agentic AI
Custom agent systems — Hosted, On-Prem, or Hybrid, run by us.
On-Prem and Hybrid first: agents inside your boundary, with routing engineered by data classification and the quality question answered by measurement, not vendor claims.
Custom agentic-AI systems — designed, built, and operated across three deployment models. Architecture, software, and infrastructure pointed at one hard thing: agents that work in production and stay working as the models underneath them move.
scripted demo · deterministic replay
the live concierge ↘ is the real thing
Who it's for
Teams that want agents doing real work in production — support and triage, document and data processing, internal automation, customer-facing assistants — without standing up an ML platform or carrying the model churn. Where the data can't leave, we build inside your boundary; where most traffic is routine but some can't go to an external API, we route deliberately.
How it fits
The agentic practice draws hardest on the other three — it is architecture, software, and infrastructure aimed at one hard problem — which is why the same team runs all four.
What's included
- Discovery and a spec for the agent system: the workloads, the integrations, the data boundary, and the evaluation plan that defines “working.”
- Design and build of the agents, tools, and integrations against your systems.
- Deployment across three models — Hosted, On-Prem, or Hybrid — chosen from your data boundaries, scale, and economics, not picked in advance.
- Evaluation harnesses so behavior is measured, not vibed — and re-measured as models drift.
- Full operation: monitoring, model upgrades and migrations, prompt and tool maintenance, and cost and performance tuning, for the life of the system.
- A documented data-handling boundary: what transits where, retained how long, under what controls.
What you provide
- A workload owner who can describe the process and adjudicate the edge cases.
- Access to the systems the agents must touch, scoped to the spec.
- Review and sign-off on the data boundary and the evaluation criteria.
Timeline shape
Discovery in weeks. The deployment model sets the rest: Hosted lands fastest, On-Prem follows hardware lead times, Hybrid goes live incrementally one workload at a time. Managed operation begins at go-live and continues for the life of the system.
Three deployment models
Hosted, On-Prem, or Hybrid — a design output, not a precondition.
Trueform Hosted
Managed agent hosting
Agents designed, deployed, and operated entirely on Trueform Agentics infrastructure, delivered as an ongoing service. The client consumes the agent system; we own everything underneath it.
Who it's for
Organizations that want agentic automation in production without standing up ML infrastructure or hiring for it. Typical buyers are mid-market companies and startups where the engineering team is fully committed to the core product, and an agent system — however valuable — would be a distraction to build and a liability to babysit. Also a fit as a first deployment for clients who expect to move on-prem or hybrid later: start hosted, prove the workload, migrate when the economics or data requirements justify it.
What's included
- Discovery and a written spec for the agent system: workloads, integrations, data flows, success criteria, and the evaluation plan.
- Design and build of the agents, tools, and integrations against your systems — APIs, data stores, ticketing, communication channels, whatever the workload touches.
- Deployment on our infrastructure, including model selection and routing: frontier APIs, hosted open-weight models, or both, as the spec dictates.
- Full operation: monitoring, alerting, incident response, evaluation as behavior drifts, model upgrades and migrations, prompt and tool maintenance, and cost and performance tuning.
- A defined data-handling boundary, documented in the spec: what client data transits our infrastructure, what is retained, for how long, and under what controls.
- Regular operational reporting: what the system did, what it cost, where it failed, and what we changed.
What you provide
- A workload owner who can describe the process being automated and adjudicate edge cases during design and evaluation.
- API access or credentials to the systems the agents must touch, scoped to what the spec requires.
- Review and sign-off on the data-handling boundary and on the evaluation criteria that define “working.”
Timeline shape
Discovery in weeks, not months. Build and initial deployment typically lands within a small number of months depending on integration surface, with the first production workload running early and scope expanding from there. Managed operation begins at go-live and continues for the life of the system.
Trueform On-Prem
On-prem design & buildout
We spec and stand up dedicated agentic-AI infrastructure inside the client's walls — from a single inference node serving one department to multi-rack GPU deployments — and either hand it over or stay on to co-manage it.
Who it's for
Organizations whose data cannot leave their boundary or whose scale makes owned infrastructure the right economics. Typical buyers are regulated and data-sovereign organizations — financial services, healthcare, defense-adjacent, public sector — and enterprises with sustained inference volume high enough that owned hardware beats per-token pricing. Also a fit for organizations with existing data-center capacity who want it earning its keep.
What's included
- Workload analysis and capacity modeling: which agent workloads run locally, what models they need, at what throughput and latency, with what growth headroom.
- Model selection and sizing: open-weight model evaluation against your actual tasks, with quantization and serving trade-offs written down, not assumed.
- Hardware specification: GPUs, CPU inference where it suffices, memory, storage, power, and cooling — sized to the workload analysis, from a modest single-node inference box up to multi-rack GPU clusters.
- Procurement support: vendor selection, quotes, and purchase coordination, with the commercial handling stated plainly up front.
- Networking and integration design: how the inference layer meets your network, identity, and security architecture.
- Buildout and commissioning: installation, serving-stack deployment, load testing, and acceptance against the spec.
- The agent systems themselves, designed and built to run on this infrastructure — the buildout is in service of working agents, not hardware for its own sake.
- A handover or co-management decision, made explicitly: full handover with runbooks and training for your team, or an ongoing managed arrangement where we operate the agent layer (and optionally the serving layer) on your hardware.
What you provide
- Data-center or server-room facilities meeting the power, cooling, and network requirements in the spec — or engagement with your colocation provider.
- Security, compliance, and network teams available during design to state the real constraints early.
- Capital budget ownership for hardware purchases.
- If full handover is chosen: an operations team to receive the runbooks and training.
Timeline shape
Longer than the other offerings, dominated by hardware lead times. Design and spec complete in weeks to a few months; procurement and delivery run on vendor schedules; buildout, commissioning, and agent deployment follow. Single-node deployments are substantially faster end to end than multi-rack ones. We sequence the agent build in parallel with procurement so the software is ready when the hardware is.
Trueform Hybrid
Hybrid architecture
Frontier-API-backed agents — Anthropic's models and others — with workload deliberately offloaded to client-local hardware at whatever tier the economics justify, and routing between them engineered by capability, cost, latency, and data sensitivity.
Who it's for
Organizations already running agents (or substantial LLM workloads) against frontier APIs whose spend, latency profile, or data exposure has outgrown an API-only architecture — but for whom full on-prem would sacrifice frontier capability they genuinely need. Typical buyers are enterprises with meaningful and growing API spend, and organizations where most traffic is routine but a subset of data or decisions cannot go to an external API.
What's included
- Workload profiling: which calls in the existing system actually need frontier capability, which could run on local models, and what each category costs today.
- Routing and escalation design: explicit, documented policy for what runs where — by task difficulty, cost ceiling, latency budget, and data classification — including when a local model escalates to a frontier API and how that escalation respects data boundaries.
- Local-tier specification: the hardware and open-weight models for the offload tier, sized to the profiled workload. This can be as small as one inference server or as large as an On-Prem-style buildout; the hybrid spec says which, and why.
- Build and deployment of the routing layer, the local serving stack, and the agent changes needed to use both.
- Evaluation harnesses that compare local and frontier output quality on your tasks, so routing thresholds rest on measurement rather than vendor claims.
- Ongoing operation of the routing layer and local tier: re-tuning as model prices and capabilities shift, as new open-weight models change the offload calculus, and as your traffic mix moves. Routing policy is revisited on a schedule, because the inputs to it do not hold still.
What you provide
- Access to current API usage data and spend, so profiling works from real traffic rather than estimates.
- Hardware for the local tier (procured through us or directly), and the facilities for it, proportionate to its size.
- A data-classification decision: which categories of data may transit frontier APIs and which must stay local. We help formalize it; you own it.
- Engineering counterparts for integration into the existing agent or application stack.
Timeline shape
Profiling and routing design complete in weeks. The local tier follows hardware lead times, which are short for single-node tiers and longer for larger ones. Routing typically goes live incrementally — one workload category at a time, each validated against the evaluation harness before the next moves — so value lands before the full architecture is in place. Operation and re-tuning continue for the life of the system.
Interactive · about 90 seconds
Not sure which model fits? Let the agent reason it out.
Answer six questions about the workload, data boundaries, scale, and priorities. The configurator — deterministic and grounded in these three models — maps them to Hosted, On-Prem, or Hybrid and reasons it out live: a rationale, a tailored architecture sketch, and the honest next step. No figures from a form.
Configure your deployment →▸ your spec
Support & triage · high volume · data can't leave our boundary
▸ reasoning
hard constraint: data-residency — nothing can leave.
→ mapping to Trueform On-Prem.
Trueform On-Prem
matchedon-prem design & buildout
Our own fleet, live
We run the firm on the same agents we build.
Seven agents — marketing, sales, support, website, ops, finance, analytics — run our business day-to-day. Every outward action launches gated to a human; gates come off, per action type, only when the action log earns it.
mid-2027
Every outward action launches gated. Gates come off per action type only when the logged evidence earns it — never wholesale. The curve is the plan in autonomy-roadmap.md, not a promise.