Architecture and memory
The 2026 AI Agent Playbook
A field brief on agent architectures and the tradeoffs in context and memory design.
Where the agent runs
These patterns illustrate different ways to host and connect agents. The right choice depends on the work and deployment constraints.
| Framework | Documented deployment pattern | Documentation |
|---|---|---|
| OpenClaw | A Gateway connects the assistant’s sessions, tools and messaging channels. | OpenClaw README |
| Hermes Agent | A terminal agent and messaging gateway with local and remote execution backends. | Hermes Agent README |
| Mastra | A TypeScript framework for agents and applications. Deploy its server independently or integrate it into an existing web application. | Mastra documentation and deployment overview |
Observation and reflection
Mastra’s Observational Memory uses observation and reflection to condense message history. An Observer produces notes from the conversation; a Reflector condenses the observation log. Read the memory documentation.
Evaluate retained facts and retrieval quality on your own conversations.
For an evaluation, keep the original conversation alongside the condensed version. Ask questions whose answers require an earlier decision or a correction. Record which details survive and which are lost.
A proposed comparison: give the same task to a system with the original context and one with the condensed context. Judge whether each follows the user’s latest instructions. This brief reports no measured compression gain or performance advantage.
A threat model to consider
Simon Willison’s threat-model discussion considers agents that can read private information while encountering untrusted content and communicating externally. Treat that combination as a design question for your system.
Mitigations to evaluate
Consider limiting access to the data a task needs and requiring review before external actions. Test whether untrusted material can change the agent’s instructions or the destination of an action. These are evaluation suggestions, not guarantees of protection.
Before publishing agent-generated material, check the rules that apply to the service and jurisdiction. The European Commission provides the official Article 50 text.
Evaluate the complete system
Evaluate the complete system on representative tasks. Record what constitutes a completed workflow and inspect failures as well as successful runs.
Measure cost and concurrency on the workload you intend to run.
Check what survives memory compression before relying on it.