AgentRecipes
Practical patterns for building real agents: memory, tools, workflow state, review loops, and learning systems.
This section is for builders who already know the model is only one part of the system. The useful questions are usually more structural: what should live in memory, what should live in state, what should be logged, and where human review should sit.
These recipes are meant to be implementation-oriented. They are not prompt tricks. They are patterns that tend to hold up better once an agent moves past the demo stage and starts doing repeated work.
What you'll find here:
- How to separate agent identity from workflow state
- How to compress session history without losing continuity
- Where to add human approval gates in real workflows
- How to keep tool-enabled agents from collapsing into prompt sprawl
- What to log if you want your system to improve from real usage
Memory
Identity vs. State
How to separate durable agent identity from changing workflow state so your agent stays consistent in real operation.
Session Summaries That Actually Help
A practical pattern for compressing conversation history into something an agent can reuse without dragging old noise into every run.
Workflow
Tooling
Learning
Need an agent system built around a real workflow?
If you already know the use case and want help turning it into something production-ready, that's the kind of work I do.