The AI agent development company for agents that pass the eval. Not just the demo.

Agent types we build

One workflow per sprint. The agent does it end to end; the eval suite says whether it did.

• Support and service agents. Ticket triage, answer drafting, refund and account actions behind approval gates, handover to a human with full context.
• Operations agents. Intake, scheduling, document handling, back-office data entry and reconciliation across your existing tools over MCP.
• Finance-decision agents. Lending, disclosure review, reporting and adverse-action explainability. Routed to QGI when a regulator can ask about the output.
• RAG and QAG over your documents. Retrieval agents on frontier models, or Chunkless RAG on the QGI QAG Engine when the answer has to cite the whole document, not a fragment.
• Workflow automation agents. Multi-step, multi-tool runs with approval gates on anything irreversible, budget caps and a versioned audit log.

Stack we ship on

Model choice is an eval result, not a preference. The suite decides, and decides again when a model is retired.

• Claude 5 family. Sonnet 5 (GA 2026-06-30), Opus 5 (2026-07-24) and Fable 5.1 (GA 2026-09-01, cache reads cut 75%) for long agentic tool loops.
• GPT-5.6 Sol, Terra, Luna. Global rollout 2026-07-09. Sol for complex reasoning, Terra for cost, Luna for volume. GPT-5.5 retires from ChatGPT and Codex on 2026-10-14.
• Gemini 3.6 Flash. Released 2026-07-21 for agents at scale; up to 17% fewer tokens than 3.5 Flash. Open-weight fallbacks (GLM-5.2, Kimi K3, Qwen 3.6) for private deployment.
• MCP, stateless spec. Your tools exposed as MCP servers on the 2026-07-28 stateless spec. Third-party servers vetted before they touch data; supply chain treated as attack surface.
• Eval suites in your CI. Offline evals on every change, online evals in production, regression gates on prompt, tool and model swaps.
• Harness with kill switch. Agent inventory, approval gates, budget caps, audit log and a kill switch you can reach. The controls 74% of teams believe they have and 19% do.

Engagement models

Three steps, each one optional after the last. No retainer required to start; no lock-in to leave.

• 1. Discovery call. Thirty minutes. You bring one workflow and what a wrong answer costs. You leave with a written eval set and a model recommendation. Free.
• 2. Scoped sprint. Fixed scope, one workflow, evals written before the first prompt. Ships the agent, the eval suite and the harness into your repo. Priced on the call against 2026 market ranges.
• 3. Run or hand over. Your infra, your keys, your kill switch. Keep us on for model swaps, drift and new workflows, or run it yourself. Either way you own the code, prompts and evals.

Industries

Two tracks, one question: does a regulator get to ask about the output?

• Regulated finance, via QGI. Banks, lenders, insurers and advisers. Deterministic decision layers built against SR 26-2, FINRA 2026 and SEC FY2026 exam language, with EU AI Act Annex III evidence for credit and insurance uses.
• SMB and mid-market operations, via TheoSym. Founders and operators who need support, ops and back-office agents in production in weeks. Frontier-model agents, evals and harness included, no compliance theatre.
• Enterprise engineering teams. Teams with agents already in pilot that need the eval suite, the inventory and the controls to graduate them. 88% of pilots never do (LangChain, 2026-03-03).

What is an AI eval?

An AI eval is a versioned set of test cases that scores an agent on real tasks, offline in CI and online in production, so a prompt change, tool change or model swap is caught before users see it. Only 19% of teams have controls that catch agent failures (Harness, 2026-09-10). TheoSym ships the eval suite with every agent.

What is MCP?

MCP, the Model Context Protocol, is the open standard that connects AI agents to tools and data through a common server interface. The spec became stateless on 2026-07-28 and the first official credential, MCPA, launched on 2026-09-14. TheoSym builds MCP-native and vets every third-party server before it touches your systems.

What is deterministic AI?

Deterministic AI is a system that returns the same output for the same inputs on every run, with a traceable path from evidence to decision. It is what model-risk teams can validate and examiners can audit. QGI builds deterministic AI for financial services with the QAG Engine, Chunkless RAG and Neural Symbolic Agents.

Which frontier models does TheoSym build on?

All three labs: the Claude 5 family (Sonnet 5, Opus 5, Fable 5.1), GPT-5.6 Sol and Terra, and Gemini 3.6 Flash, plus open-weight fallbacks such as GLM-5.2, Kimi K3 and Qwen 3.6 for private deployment. The eval suite decides which model runs, and decides again when a model is retired.

When is QGI the right track instead of TheoSym?

When the output is a financial decision a regulator can ask about. QGI (Quantum General Intelligence Inc., San Diego) is an independent research company led by Dr. Sam Sammane, Founder & CEO, and Waseem Sammane, VP Engineering. TheoSym is one of its implementation partners; the same form routes you to either track.

What does an AI agent cost in 2026?

Published agency ranges: $5K to $25K for a single-purpose agent, $25K to $100K for multi-agent systems with orchestration, and $250K to $400K+ for enterprise platforms with compliance. Run costs land at $200 to $2,500 per month. These are market ranges, not TheoSym prices; your scope sets your number.

How fast can an AI agent development company ship?

Published MVP ranges sit at 6 to 10 weeks and buyers now expect weeks, not quarters. TheoSym scopes one workflow first and writes the evals before the first prompt, so the schedule is set by the pass rate, not by a slide. You get a date on the strategy call.