When should software choose the next step?
Planned comparison: a fixed workflow versus bounded tool use. No comparative results yet. Measure completion, incorrect actions, recovery and cost.
Research & engineering notes
Explore the possibility.
Test the useful part.
A public record of what informs our engineering, what IMFIT demonstrates and what we want to investigate next.
Research agenda / proposed experiments
Questions worth
building to answer.
Our next capabilities should earn their place through useful tasks and measured results. These are research directions, not completed products or published benchmarks.
Can a recommendation explain its sources?
Proposed study: turn an unstructured request into a shortlist with explicit constraints and current records. No recommendation product or measured results yet.
How much model does the task need?
Use case not selected. Compare rules, hosted models and small models once a real task and evaluation data are available. No fine-tuned model to report.
No custom small-model training, autonomous agent deployment or recommendation product is claimed here.
Built: approved knowledge in IMFIT.
Review, versioning and response context from our owner-operated implementation. The note distinguishes inspected code and selected local checks from broader evaluation still to do.
Read the implementation noteResearch: agents or a defined workflow?
Choose autonomy only where selecting the next step adds value. Our IMFIT evidence covers AI-assisted workflows; it does not establish a general autonomous agent.
Read the architecture guideResearch: the right connection.
What should an AI application be allowed to read or do? Start with the task, permissions and available interfaces.
Explore APIs, retrieval and MCPBefore an experiment becomes a project.
Name the user, compare existing tools and agree on evidence that would justify further development.
Use the project checklistThe questions
behind the research.
Public sources establish available capabilities, not demand for Anvayric. Our conclusions remain hypotheses until tested with real tasks and buyers.
- Tool use is available.
- Anthropic distinguishes fixed workflows from agents that choose their own steps. We use that distinction to ask whether added autonomy is necessary. Read the architecture guidance.
- Small models are an option.
- Microsoft documents Phi customisation and local deployment. That is a starting point for comparison, not evidence that a smaller model will be accurate or cheaper for a particular task. Read about Phi.
- Integration includes the interface.
- MCP Apps supports interactive interfaces in compatible clients. A useful experiment should test both tool behaviour and how a person reviews its result. Read the MCP Apps introduction.
Research agenda recorded 30 October 2026. No benchmark results or customer demand are implied.
Scoped AI engineering
Have a task in mind?
Let’s assess the fit.
Start with the work, the information it needs and the question you want to answer.