48 lines
6.1 KiB
Markdown
48 lines
6.1 KiB
Markdown
# Harness
|
||
|
||
#concept
|
||
|
||
## Summary
|
||
|
||
A **harness** is a universal agent = LLM + a small standard toolset + a runtime loop. The metaphor (Konstantin): the LLM is the force, the tools are the harness/upness, the data is the field — the LLM "pulls the tools across the data space," turning raw tasks into solved ones. Different tasks, *same* processing method.
|
||
|
||
## Current Understanding
|
||
|
||
The harness is the de-facto unit of agentic work in 2026. A good one has: a **short** system prompt ("use the tools to solve the task"), ~10–40 standard tools (Read/Edit/Write, Bash, Grep/Glob/WebSearch, TodoWrite, subagent spawn), a ReAct runtime loop, context management (timely summarization, subagents without full context), standardized key files (CLAUDE.md/AGENTS.md, `skills/`), resilience to failures, and an interface (CLI/chat/messenger). It runs in two modes: **interactive** (human-in-the-middle → man-on-the-middle) and **autonomous** fire-and-forget (CLI + task arg, embeddable in CI/CD).
|
||
|
||
**Tool ceiling:** >30–40 built-in tools appears in no popular harness; >100 confuses even Fable 5 because every tool is injected into each call — a key reason [[skills-as-memory|skills]] (two-stage loading) beat raw tools. See [[context-as-scarce-resource]].
|
||
|
||
**Consolidation over tool-hopping.** The [[2026-07-21-larysa-interview|Larysa interview]] sharpens Eugene's pitch into its user-facing form: assemble *one* place where all your project agents live and can talk to each other, instead of spinning up the tool in separate folders and juggling apps. The claimed payoff is as much cognitive as productive — "you build yourself an operating system where you have everything in one place… When I built this thing, my life just split into before and after" — freeing mental overhead he hadn't noticed he was spending. Completion notifications ("done thinking") are part of the claim: they are what make background agents usable without babysitting. This is the engineer's phrasing of the webinar's title arc, chat box → your own OS ([[levels-of-ai-usage]], [[personal-ai-operating-system]]).
|
||
|
||
**The business-facing formula.** An anonymous Russian business short ([[2026-07-22-ai-is-stupid]]) independently restates the concept for non-engineers: the harness is an "engineering wrapper" — what the model must verify, which tools to trust, how to shape the answer, what is forbidden — and **strong model + your business context + harness = employee-level answer**. Remove any component and you get "smart but generic," "specific but undisciplined," or "stupid AI." Useful as webinar language: it names what the audience already feels (generic answers) without requiring the engineering vocabulary.
|
||
|
||
**A second reference harness: [[amp]].** Where [[claude-code]] is local-first and user-extended, AMP is sandbox-first and vendor-curated ([[2026-07-28-agentic-engineering-10x-developer]]): a PWA install, a low/medium/high/ultra dial that maps each level to a model *and* a sub-agent set, named sub-agents (**Oracle** the reviewer, **Painter** the image generator), a meta-agent (**Puck**) that spawns and messages other agents, and **orbs** — remote sandboxes where one URL carries thread + agent + computation + diff (see [[async-by-default]]). Two things it demonstrates about the concept: the harness is now the *product* (AMP's most-asked customer question is "what's the meta — what model, what prompt?", i.e. customers pay for research decisions), and a harness can be strong with **no user-authored skills layer at all** — the structure exists, but the vendor supplies it. That is the design axis [[hermes]] and Claude Code put in the user's hands.
|
||
|
||
**The governance fault line:** [[eugene]] argues every developer should **build their own** harness (deep knowledge → more effective). [[sebastian]] counters that "bring your own harness" cannot survive enterprise compliance — it must be a company-managed resource, and *that gap is the business*. See [[enterprise-ai-reality]].
|
||
|
||
## Evidence
|
||
|
||
- Definition, toolset, two modes, popular harnesses (Claude Code, Codex CLI, Cursor, OpenClaude, Hermes) — [[2026-07-14-skills-based-on-git]].
|
||
- BYO-harness demo and the enterprise counter-argument — [[2026-07-14-sebastian-eugene-interview]].
|
||
- Orchestration-era models that "spawn sub-models and verify" are the model-side of the same idea — [[2026-07-14-everything-we-knew-about-software-has-changed]].
|
||
- Claude Code as the reference harness across surfaces — [[2026-07-14-gap-between-ai-users-irreversible]].
|
||
- Consolidated multi-project workspace, inter-agent messaging, completion signals, "before and after" claim — [[2026-07-21-larysa-interview]].
|
||
- Harness as "engineering wrapper"; model + context + harness formula; "stupid AI" as the harness-less default — [[2026-07-22-ai-is-stupid]].
|
||
- AMP's dial/sub-agents/meta-agent/orbs; "what's the meta?" as the customers' recurring question — [[2026-07-28-agentic-engineering-10x-developer]].
|
||
|
||
## Related Pages
|
||
|
||
- Tools: [[claude-code]], [[hermes]], [[amp]]
|
||
- Concepts: [[skills-as-memory]], [[agentic-loops]], [[evolution-of-agent-tooling]], [[context-as-scarce-resource]], [[enterprise-ai-reality]], [[personal-ai-operating-system]], [[leave-less-room-for-imagination]], [[async-by-default]], [[shedding-weight]]
|
||
- Entities: [[eugene]], [[sebastian]], [[konstantin]], [[larysa]], [[thorsten-ball]]
|
||
|
||
## Contradictions / Uncertainty
|
||
|
||
- Personal vs company-managed harness is an unresolved tension (Eugene vs Sebastian), not a settled answer. [[amp]] adds a third option neither of them argues for: a **vendor-managed** harness, where the research decisions are the purchase and the user tunes almost nothing.
|
||
- **Local vs remote.** Eugene's consolidation pitch assumes one place *on your machine*; Thorsten predicts local dev disappears into remote sandboxes. Compatible only if the thing being consolidated is the interface rather than the compute. Status: tentative.
|
||
- The "life split into before and after" consolidation payoff is self-reported by its builder and never measured; Larysa, the practitioner it was pitched to, does not yet run one. Status: tentative.
|
||
|
||
## Next Questions
|
||
|
||
- What is the minimal compliant, centrally-managed harness a regulated team could adopt?
|