An AI model can describe how to investigate a problem. An agent system can work through it: choose an action, inspect the result and decide what to do next. The useful unit is the whole system: model, harness, tools, environment and feedback.1
This collection connects the mechanisms with the systems people actually use. Choose a route first; the chapters supply the depth.
A map of the system
The model proposes decisions. The harness is the surrounding software that assembles context, manages the loop and coordinates execution. Persistent session records and execution environments can have different lifecycles; they need not live inside one process.23 Figure 1 connects that work loop to the wider landscape: packaged products, extensible harnesses, frameworks and managed runtimes. Select a label to explore the corresponding chapter or study.
- ContextGoal, instructions and relevant evidence.
- ModelChoose an action or response.
- ActionTools act in files, apps or services.
- EvidenceObserve results; check the outcome.
- Coding products
Package an end-user workflow.
- Extensible harnesses
Provide reusable agent machinery.
- Frameworks & SDKs
Build an application-owned system.
- Managed runtimes
Operate harness and session infrastructure.
Roles overlap; examples are not a quality ranking. DeepSeek Harness is a developer preview.
Figure 1. Latent Light synthesis: the work loop, overlapping ecosystem roles and frontier questions. Implementation and research sources are collected in the deep-dive comparison.1234
A fixed workflow chooses its path in advance. An agent delegates more path selection to the model. Many useful systems combine both: deterministic steps handle known work, and a model handles uncertainty.5 More freedom is useful only when the system can observe and verify what it does.
Learn the mechanisms
| Question | Chapter |
|---|---|
| How does a response become a sequence of actions? | How Agents Work |
| What can the system do, and how does it connect to other components? | Tools & Skills |
| What is visible now, what persists, and how does work resume? | Context & Memory |
| When should control live in code, one agent or several workers? | Workflows & Coordination |
| What does the whole process look like on a repository change? | A Coding Workflow |
| How do we know the outcome is correct and repeatable? | Evaluation |
| How do we bound access, actions and external effects? | Security & Control |
For a first reading, follow the first three chapters. Before designing a system that changes external state, read evaluation and control as well. The coding workflow connects those ideas in one worked example.
Study the systems
The deep-dive overview compares the five initial studies by execution ownership, persistence, extension mechanisms and control. It also connects them to the broader ecosystem, including frameworks, SDKs and managed runtimes.
The studies ask different questions. What does Claude Code package into its workflow? How does Codex expose a controllable session to clients? What does Pi leave to its host and extensions? What does OpenCode’s client/server boundary make possible? What changes when DeepSeek makes the loop itself a plugin? Follow a system from its design to its consequences, rather than counting features.
Follow the frontier
The frontier discussion examines durable sessions, evolving context, coordinated workers and feedback that can genuinely grade an outcome. Research findings and vendor engineering accounts have different limits; a documented feature does not establish reliable task completion.
The individual studies identify their inspected documentation or repository snapshot. Source checks for the five studies are October 11, 2026. The broader directory retains its separately dated checks. These pages explain implementations and research; they do not report a new head-to-head benchmark.
References
-
Shunyu Yao et al. ReAct: Synergizing Reasoning and Acting in Language Models, ICLR 2023. ↩ ↩2
-
Anthropic. Scaling Managed Agents: Decoupling the brain from the hands, April 8, 2026. Session, harness and sandbox responsibilities; checked October 11, 2026. ↩ ↩2
-
OpenAI. Agents API overview, managed harness and application integration; checked October 11, 2026. ↩ ↩2
-
Latent Light. Agent Deep Dives, synthesis of the five inspected implementation studies, ecosystem roles and bounded frontier evidence; source checks October 11, 2026. Categories overlap and do not indicate measured quality. ↩
-
Anthropic. Building effective agents, 2024; workflows and agents. ↩