Study an agent by following one task across its boundaries: what the model sees, how actions run, what persists and who verifies the result. The five initial deep dives expose different design choices inside that common problem.
Explore the studies · Compare responsibilities · Follow the frontier
The five studies
These are architectural studies of selected systems, not a popularity ranking. Current documentation and the three open-source snapshots were inspected October 11, 2026. Their worked scenarios illustrate mechanisms; they do not report an executed head-to-head benchmark.12345 DeepSeek Harness remains an explicitly unaudited developer preview.6
Put each system on the landscape
A coding product, reusable harness, framework and managed runtime can overlap. Figure 1 shows the responsibility emphasized by each category. A project can occupy several places on this map.
- Coding products
Package a work loop into a user-facing development workflow.
ExamplesClaude Code, Codex, OpenCode
Question to testWhere does work run, and how does the user inspect and steer it?
- Extensible harnesses
Assemble reusable context, tool and execution machinery.
ExamplesPi, DeepSeek Harness, Deep Agents
Question to testWhat does the host own, and which components can it replace?
- Frameworks and SDKs
Provide primitives for constructing an application-owned system.
ExamplesLangGraph, Google ADK, agent SDKs
Question to testWhich persistence, retry and policy responsibilities remain in the application?
- Managed runtimes
Operate parts of the harness and session infrastructure as a service.
ExamplesOpenAI Agents API, Anthropic Managed Agents
Question to testWhich control, storage and execution boundaries does the service actually own?
Figure 1. Latent Light synthesis of product, harness, framework and managed-runtime responsibilities.123457891011
The broader ecosystem directory retains additional study candidates, including Microsoft Agent Framework, smolagents and Strands. Its product details retain the source-check dates stated there. The Personal AI Agents landscape examines consumer delegation and user needs; this collection examines the mechanisms and implementation responsibilities.
Compare the system boundaries
Read across the table to compare responsibilities, then open a study for its source details and qualifications. Interface names alone cannot establish execution location or isolation.
| Boundary | Claude Code | Codex | Pi | OpenCode | DeepSeek Harness |
|---|---|---|---|---|---|
| Study focus | Project-aware coding workflow | Controllable session and execution lifecycle | Reusable loop and extensible host | Client/server coding runtime | Replaceable plugin composition |
| Integration unit | Product workflow and extensions | Interactive clients, SDK and App Server | In-process SDK or subprocess RPC | Server API and clients | Profile, bundles and service plugins |
| Context and continuity | Guidance, memory, compaction and saved sessions | Threads, work items and configured context management | Active branch of a persistent session tree | Server-owned sessions and child sessions | Derived history from a durable session log |
| Extension question | Which skill, hook, tool or worker fits? | Which instructions, integration and client contract fit? | Which executable extension or resource belongs in the host? | Which plugin, skill or configured role fits? | Which service or lifecycle seam should change? |
| Control question | Inspect permission and execution settings | Separate approval from filesystem/network containment | Supply external containment; no built-in resource permission system | Inspect allow/ask/deny rules and host access | Inspect the mounted guards and preview limitations |
This table is our comparison lens, informed by the detailed contracts cited in the individual studies. It does not score reliability or security. “Supports memory” is particularly weak as a comparison: active context, a recoverable session record and remembered knowledge solve different problems. LangGraph’s checkpointer/store distinction offers another concrete example of these scopes.8
A meaningful trial would hold the task, model where possible, environment and total budget fixed, then inspect outcomes, recovery and cost. If models or scaffolds differ, report that difference instead of attributing the entire result to the product name. See Evaluation.
Frontier
Several active directions are changing what a useful harness needs to provide. The evidence below mixes inspected research with labeled vendor engineering accounts. Each supplies a bounded lesson.
Durable sessions beyond one running process
Anthropic’s managed architecture separates the session log, harness and sandbox. The record can survive a failed harness, while the next harness reconstructs work from events. Its account explicitly distinguishes the full session from Claude’s selected context window.10 OpenAI’s Agents API similarly assigns managed harness and session responsibilities to the service, with environment choices remaining explicit.11
The practical question is what can be recovered after an interruption. Persisted text is useful, but an external operation may have committed before its response was lost. Durable execution must reconcile effects as well as reconstruct a conversation. These vendor designs do not prove that every application using them already implements that recovery contract.
Evolving context with accountable updates
ACE studies context adaptation without changing model weights. It separates generation, reflection and curation, then merges incremental context updates. Figure 2 shows the authors’ mechanism. The reported evaluations concern particular tasks and baselines; they do not establish universal improvement from accumulating more memory.12
Figure 2. ACE context adaptation, reproduced from Zhang et al., version 3, original Figure 4, under CC BY 4.0. Converted to WebP without changing the diagram.12
The important design question is who may revise stored guidance, with what evidence and rollback. An incremental memory entry can preserve a useful lesson or preserve an error. The paper’s mechanism motivates explicit review of context changes rather than treating a longer prompt as a guarantee.
Coordination that fits the task
Kim and colleagues compare agent architectures across multiple benchmark, model and configuration settings. Their reported results vary with task structure: decomposable work and sequential planning do not benefit equally from coordination. The paper identifies overhead and error propagation as consequential factors.13
That supports a conditional design rule: add workers to resolve a particular failure or independent work split, then measure the whole job. It does not establish that all multi-agent systems help or fail. The coordination chapter explains the contracts and costs.
Feedback that can actually distinguish improvement
Anthropic’s application-development account uses an evaluator that interacts with the running application and is calibrated against explicit criteria. It also reports much greater cost in one multi-agent comparison and describes simplifying the harness as the generator model improved.14
This is a vendor engineering account with selected examples, not a controlled ranking of these five products. Its useful lesson is that a critic needs observable evidence and a discriminating rubric. A more capable model can also change which scaffolding earns its overhead. Repeated self-approval is not verification.
Choose the next question
| If you want to understand… | Start with… |
|---|---|
| A packaged development workflow | Claude Code |
| An application-controlled coding session | Codex |
| A host you can adapt through a small core | Pi |
| Multiple interfaces around a runtime API | OpenCode |
| Replaceable services and lifecycle composition | DeepSeek Harness |
These are study routes, not adoption recommendations. For design prerequisites, return to How Agents Work; for the practical end-to-end example, read A Coding Workflow.
References
-
Anthropic. How Claude Code works, documented harness and work loop; checked October 11, 2026. ↩ ↩2
-
OpenAI. Codex SDK, local automation and App Server distinction; checked October 11, 2026. ↩ ↩2
-
Mario Zechner and Earendil Works. Pi, toolkit and host permission boundary. Snapshot checked October 11, 2026. ↩ ↩2
-
OpenCode contributors. Server, client/server API. Snapshot checked October 11, 2026. ↩ ↩2
-
DeepSeek AI. Architecture, plugin composition, profiles and services. Snapshot checked October 11, 2026. ↩ ↩2
-
DeepSeek AI. Safety, experimental status and absence of a security audit. Snapshot checked October 11, 2026. ↩
-
LangChain. Deep Agents overview, reusable harness; checked October 11, 2026. ↩
-
LangChain. LangGraph persistence, checkpointers and stores. Checked October 11, 2026. ↩ ↩2
-
Google. Agent Development Kit, agent-development framework. Checked October 11, 2026. ↩
-
Anthropic. Scaling Managed Agents: Decoupling the brain from the hands, April 8, 2026; session, harness and sandbox separation. Checked October 11, 2026. ↩ ↩2
-
OpenAI. Agents API overview, managed Codex harness and environment choices; checked October 11, 2026. ↩ ↩2
-
Qizheng Zhang et al. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models, ICLR 2026; version 3, March 29, 2026, Sections 2–3 and Figure 4. CC BY 4.0. ↩ ↩2
-
Yubin Kim et al. Towards a Science of Scaling Agent Systems, version 3, April 8, 2026; abstract and reported benchmark scope. Preprint. ↩
-
Prithvi Rajasekaran, Anthropic. Harness design for long-running application development, March 24, 2026; evaluator calibration, cost and model-dependent simplification. Engineering account, checked October 11, 2026. ↩
