Interactive agent systems observatory

An AI agent, with every boundary exposed

An agent is not just a model with a personality. It is a model inside an engineered system: finite context, callable tools, persistent state, a control loop, trust boundaries, budgets, policies, and a definition of done.

Change one mechanism at a time. Every count comes from the visible teaching graph. Security controls are labeled as soft model signals or hard system boundaries. No intelligence scores, no hidden chain-of-thought, and no decorative autonomy theater.

Agent systems observatory · Every runtime boundary stays visible

Six experiments: Context · injection · loops · extensions · topology · real runtimes
Observable state: Tokens · calls · messages · policies · stop reasons
Research date: Official sources checked August 10, 2026

Shared teaching caseResearch six AI conferences → verify dates → create a sourced briefing → save a draft → send only after approvalEvery experiment exposes a different system layer around the same kind of work.

Experiment 01

Context engineering

contextₜ = select(instructions, tools, evidence, memory, history)

Context compiler3,000 / 3,000 fixed teaching tokens packed

Available state

Model context at turn t100% full

System policy310 t

User goal190 t

Tool schemas640 t

Activated skill520 t

Six event sources980 t

Durable memory360 t

Selected6 blocks

Excluded1 blocks

Model weightsunchanged

Recompiledevery turn

Context budget3,000 teaching tokens

Prompt writing is one input. Context engineering is the whole selection system.

It decides which instructions, tool descriptions, evidence, memories, and recent messages enter a finite window—and what stays out.

More context can be worse context.

The packer uses transparent priority rules, not a claim that token count predicts quality. Toggle the raw dump and watch it displace more targeted state.

Why “teaching tokens”?

The fixed allocations make displacement arithmetic reproducible. They are not presented as counts from a particular provider tokenizer; real counts vary by model and serialization.

Trusted: Application or user instructions with known authority
Untrusted: Content to inspect, never authority to inherit
Generated: Agent-created state that still requires provenance

Primary research

Inspect the mechanism at its source.

Project architecture and terminology were checked against official documentation on August 10, 2026. Research papers are linked to their original publication records.

Agent and workflow patternsBuilding effective agentsOriginal source ↗Tools, resources, and promptsModel Context ProtocolOriginal source ↗Portable procedural contextAgent Skills specificationOriginal source ↗Gateway and execution systemOpenClaw architectureOriginal source ↗AIAgent and learning systemHermes Agent architectureOriginal source ↗

Keep experimenting

Change the layer, keep the system visible.

AI Lab indexSee every interactive laboratoryBrowse the suite →Transformer microscopeGo beneath the agent runtime into the modelInspect the matrices →