04 — Current landscape: compare architectures, not brands

A feature list expires; an architectural reading lets you understand the next product that appears. This unit uses fifteen systems reviewed on August 23, 2026 to teach you how to locate who owns the loop, where context comes from, where authority is enforced, and what kind of result can be attributed to the harness.

The selection is a reference set chosen for architectural diversity, not a census or ranking. The absence of a product implies no quality judgment. The structured catalog and its sources are in harness_landscape.json.

1. Five layers commonly called a “harness”#

Layer	What it owns	Engineering question	Catalog examples
Coding agent runtime	One agent's loop, repository context, tools, and session	How does it turn model decisions into verifiable changes?	dsh, OpenCode, Goose, Crush, Aider, Claude Code, Pi, echoVic's Orca
Personal agent runtime	Persistent memory, skills, channels, and subagents	What does it learn across sessions, and what must it forget?	Hermes Agent, OpenHarness/Ohmo
Integrated agent product	Runtime plus IDE context, CLI, and approval UX	Which advantage comes from the editor and which from the loop?	Kilo Code, Cline, Continue
Multi-agent ADE	Processes, worktrees, and supervision for multiple agents	How does it isolate, compare, and merge candidates?	stablyai's Orca
Workflow orchestrator	Deterministic stages that delegate agentic turns	Which control is expressed in code instead of requested from the model?	VirtusLab's Orca

The layer is not a score. An orchestrator is not “more advanced” than a runtime: it solves a different problem. You may attribute a difference to the harness mechanism only when both candidates occupy the same layer, receive the same surface, and work under equivalent contracts.

2. Fifteen systems, fifteen useful questions#

System	Architectural property that matters	Worthwhile experiment
DeepSeek Harness (`dsh`)	Adapter, tools, log, sandbox, and even the agent loop are plugins	Replace only the loop and measure recovery without changing model or authority
OpenCode	Coding runtime with terminal, desktop, IDE, and headless surfaces	Fix one surface and put context and compaction under pressure
Goose	General-purpose agent extended through MCP, not restricted to code	Compare a minimum and full manifest with the same model
Crush	Native TUI with sessions and compatible endpoints	Separate process overhead from loop quality
Aider	The model emits textual edit protocols, not the same tool-calling mechanics	Use it as a baseline for models that fail structured calls
Hermes Agent	Memory, skills, and subagents survive an isolated task	Measure legitimate transfer on a second pass without leaking the answer
Claude Code	Reference runtime with a demanding capability surface	Measure schema pressure and recovery without assuming undocumented compatibility
Pi	Extensible runtime with four default tools and no built-in permissions	Isolate it externally and sweep tool-surface size
Orca — stablyai	Coordinates existing agents in parallel worktrees	Measure best-of-N, duplicated cost, contamination, and merge load
Orca — VirtusLab	Typed, deterministic, resumable staged workflows	Assign a runtime/model per role and measure the composed system
Orca — echoVic	Third-party Rust coding agent with resumable sessions	Verify whether its specific defaults help under a fixed endpoint and sandbox
OpenHarness/Ohmo	Tools, skills, memory, and coordination infrastructure plus a reference agent	Separate framework value from Ohmo's behavior
Kilo Code	One product across VS Code, JetBrains, and CLI	Compare IDE and CLI while controlling diagnostics and ambient context
Cline	Shared engine across IDE, CLI, and SDK with human approval	Keep the same approval regime across every trial
Continue	Combines agent workflows with editor assistance and autocomplete	Evaluate each workload separately, never with one score

The catalog's final columns separate verified facts from hypotheses. “It has memory” may be a design fact. “Memory improves the result” remains a hypothesis until an experiment controls contamination, the first pass, and the second.

3. The Orca case: one name, three identities#

Treating “Orca” as one tool destroys any comparison:

stablyai/orca is an agent development environment. It launches other coding agents and separates their Git state through worktrees. The loop still belongs to Codex, Claude Code, OpenCode, Pi, or another delegated process.
VirtusLab/orca expresses development workflows as Scala stages. Stages are resumable, while planning, implementation, and review roles delegate to configurable agent backends.
echoVic/orca-agent is a terminal coding agent. It is a third-party project; “DeepSeek-native” does not mean it belongs to DeepSeek.

Therefore, Pi versus stablyai's Orca does not answer “which harness is better.” Valid comparisons are Pi versus another runtime under the same task contract, or Pi inside Orca versus Pi without that orchestration to measure isolation and overhead. Lab 03 rejects the first question and accepts the second as a composition test.

4. Five design theses that can be tested#

Deep replaceability#

The official dsh architecture makes the model adapter, tool registry, session log, and loop plugins. It also derives model-visible history from a durable event stream. This creates a valuable experimental seam: modifying a loop without forking the rest of the product. The tradeoff is that the effective configuration, not the binary's name, defines the harness that actually ran.

Minimum surface#

Pi provides read, write, edit, and bash by default, allows the selection to change, and keeps compaction, tools, and extensions open. This austerity reduces initial schemas but does not create a security boundary: its own documentation says the process inherits access to filesystem, network, credentials, and processes. Prompt minimalism and least privilege are different properties.

Text protocol as a baseline#

Aider maintains explicit edit-format families such as whole file, search/replace blocks, and unified diff. You must not score a response as a “failed tool call” when it never used that protocol. You can compare the outcome of the same repair, but the conclusion applies to the complete system and must retain tokens, application attempts, and patch conflicts.

Persistence as an experimental treatment#

Hermes and OpenHarness include memory or learning that exceeds one run. To measure it, the control condition must execute two sessions without memory and the treated condition two sessions under a declared retention policy. The second task must be analogous, not identical; otherwise the test measures solution memorization. Memory entries belong to the auditable dataset.

Orchestration above the loop#

stablyai's Orca proposes parallel candidates in worktrees; VirtusLab's turns planning, implementation, and review into deterministic, resumable stages. Neither magically improves a model call. They may improve the outcome through selection, specialization, or recovery while paying more compute, coordination, and effect surface. Those costs belong in the result.

5. Ambient context is also a variable#

An IDE may provide open files, selection, LSP diagnostics, and editor state that a TUI cannot see. A persistent app may remember conversations that a headless CLI does not receive. An ADE may offer five candidates while the baseline runs only one.

Before comparing, record:

exact surface: CLI, TUI, IDE, app, API, or headless;
initial context and discovery rules;
visible tools, complete schemas, and approval policy;
filesystem, worktree, variables, credentials, and network policy;
memory, caches, history, and state that cross session boundaries;
who compacts, when, and what evidence is discarded;
human process: extra prompts, approvals, selection, and merge.

If you cannot equalize a variable, do not hide the difference. Declare it as a confounder and limit the claim to “this complete system obtained this outcome under these conditions.”

6. Security: where marketing ends#

A worktree separates Git branches and directories; by itself it does not separate network, keychain, processes, environment variables, or shared services. An approval dialog may improve supervision; it does not replace permission checks in the executor. A compatible endpoint may accept messages; it does not prove equivalent tool, streaming, cache, or error semantics to the original provider.

For harness adoption, trace at least these boundaries:

which code is installed and which installation scripts execute;
where sessions, prompts, traces, memory, and credentials persist;
which telemetry or external network is used in minimum configuration;
which identity creates processes and accesses the filesystem;
which actions bypass approval and who can change that policy;
how a session is revoked, state is cleaned, and cleanup is proven.

“Open source,” “local,” and “sandboxed” answer different questions. Verify all three separately.

7. How to keep the landscape alive#

An acceptable catalog update needs a primary source, review date, and concrete architectural difference. Do not accept a card copied from a landing page or a performance claim without trials.

The pull request must:

link the official repository and documentation;
assign a layer, cohort, and composition seam;
separate verifiable facts from benchmark hypotheses;
declare surface, loop, protocol, portability, and trust boundary;
add a case that breaks if namesake products are merged;
avoid stars, maturity adjectives, and ephemeral models as quality evidence.

Codex CLI, Gemini CLI, Qwen Code, and new runtimes are obvious candidates for expanding the set, but they must pass through the same process. SOTA does not mean pretending to be exhaustive: it means the method detects a new architecture, updates the evidence, and does not preserve stale claims as truth.