AGENCY NEEDS CONTROL
The useful question is not whether a model can act. It is whether the surrounding system can constrain, evaluate, explain and recover from that action.
JASON//SYSTEM / AI ROADMAP
BUILDING AI
THAT CAN
BE OPERATED.
The current roadmap is less about adding a chat box and more about engineering useful systems: grounded retrieval, tool use, durable state, evaluation, authorization and explicit human control.
UPDATED AUGUST 2026 / ENGINEERING DIRECTION / CAPABILITY STATUS VARIES BY PRODUCT
CURRENT SIGNAL
01
BUILDING
- Agentic product workflowsMulti-step software in which a model selects tools inside explicit instructions, permissions and stop conditions.
- Retrieval-grounded experiencesProduct answers connected to approved source material, metadata, provenance and freshness.
- Operator control planesApprovals, queues, state, retries, audit logs and recovery made visible to the person responsible for the system.
02
EXPLORING
- Context engineeringDesigning the instructions, retrieved evidence, memory and tool results available at each decision.
- Model Context Protocol (MCP)A standard boundary for exposing approved tools, resources and prompts to AI applications.
- Multimodal workflowsSystems that can combine text, images, audio and computer interaction when the use case needs them.
- Single-agent firstAdding specialized agents only when simpler orchestration stops being clear or reliable.
03
LEARNING
- Retrieval-Augmented Generation (RAG)Hybrid retrieval, embeddings, vector search, reranking and citations that ground model output in controlled sources.
- Evaluations (evals) + tracingRepeatable tests and execution traces that measure quality, safety, latency, cost and tool behavior.
- Structured outputsSchema-constrained model results that application code can validate and use safely.
- Human-in-the-loopRequired review or approval before sensitive, irreversible or high-stakes actions.
04
QUESTIONING
- Where should autonomy stop?Useful automation still needs explicit authority limits.
- How much state should an agent carry?More memory can increase both leverage and error propagation.
- What becomes more valuable when code gets cheaper?Architecture, judgment, verification and domain understanding are strong candidates.
05
WATCHING
- Human + AI software developmentThe development process itself is becoming a collaboration interface.
- Machine-operated workflowsWhere continuous software can remove waiting without removing accountability.
- The economics of AI-built softwareWhat happens when implementation gets cheaper but reliable operation still costs attention.
AI SYSTEMS ROADMAP / 2026
FROM MODEL
TO OPERATING
SYSTEM.
The model is one component.
Reliable outcomes come from the system around it.
01 / GROUNDCURRENT FOUNDATION
Ground every answer.
Use Retrieval-Augmented Generation (RAG), hybrid search, embeddings, vector stores, reranking, citations and source freshness so the system can show what informed its response.
02 / ACTACTIVE BUILD
Give tools narrow contracts.
Use tool calling, Model Context Protocol (MCP), structured outputs and explicit instructions to let an agent complete bounded work without receiving unlimited authority.
03 / VERIFYCONTINUOUS PRACTICE
Measure before promotion.
Use evaluations (evals), trace review, guardrails, approval thresholds and adversarial cases before expanding a workflow's scope or autonomy.
04 / OPERATEOPERATING STANDARD
Design for the operator.
Make long-term memory, durable state, permissions, retries, audit logs, failure recovery and human-in-the-loop checkpoints part of the product—not hidden infrastructure.
This roadmap describes Jason's engineering direction. It does not imply that every capability is active in every product.
CURRENT QUESTION
If software can increasingly decide what action comes next, what should the human remain responsible for?