Towards a harness that can do anything
Developers are debating how best to “harness” large language models: whether via generic, Unix-style, file-centric frameworks like the proposed Ambiance system, or through tightly scoped, domain-specific orchestrators that wrap tools and tests around an LLM. A recurring theme is pushing as much work as possible into deterministic code and scripts—using LLMs only for judgment calls and edge cases—while grappling with practical issues like token costs, reliability, auditability, and platform constraints. Many see future AI workflows resembling composable Unix tools or OS-level services, with LLMs lightly embedded rather than acting as autonomous, free-roaming agents.
What a “harness” is and why people care
- A harness is described as the glue between an LLM and tools: it parses tool calls, runs commands (FS, network, etc.), and feeds results back.
- Some see it as an “agentic exoskeleton” or deterministic scaffold around an otherwise fuzzy model.
- Others mock the terminology as buzzwordy, but accept the underlying need: constrain and structure LLM behavior.
Deterministic workflows vs “pure” agents
- Strong current in favor of determinism: use real code and scripts for repeatable workflows; invoke LLMs only for judgment/edge cases.
- Patterns mentioned: decision trees where most leaves are “run script,” a few are “ask LLM”; code+tests as the skeleton, LLM as occasional helper.
- People describe wrapping tools like Claude Code/Codex in outer deterministic loops (tests, git, security checks, pre-commit hooks).
- Some frameworks (langgraph, ACP, etc.) are cited as good for orchestrating these loops, though a few dislike systems that make “agent-building” feel ritualized.
Unix philosophy, “everything is a file,” and alternatives
- Many like mapping agent concepts to Unix primitives: event-driven workflows, FS as shared state, FUSE, “agent as Linux user” with permissions and mail.
- Others reject “everything is a file” for LLMs, arguing that for models “everything is tokens/embeddings” and vector DBs or typed JSON are more natural.
- Files are seen by some as a pragmatic local minimum: good for both humans and models; others call FHS dated and suggest Nix/Plan 9–like ideas.
Generic vs domain-specific harnesses
- Several argue domain-specific harnesses (e.g., for software engineering) already beat generic ones: built-in ADRs, planning, behavioral tests, strict gating.
- Running LLMs “raw” is compared to driving without brakes; others report abandoning complex harnesses and going back to simpler CLIs because the extra complexity didn’t help.
Costs, tokens, and practicality
- Concern about token costs for FS-watching or frequent polling; suggestions to trigger on meaningful file thresholds instead.
- Some emphasize tests, logging, and minimal, non-overengineered scripts; overgeneralization is seen as a maintenance trap.
Reception of Ambiance and the “do anything” claim
- Positive: Unix-native mental model, lean/auditable approach, a good kernel to build variants on.
- Critiques: “soft” ideas, vague about concrete gains vs existing sandboxes; macOS-only support annoys some; “can do anything” branding viewed as overclaim.