ACI
ACI (AI Computer Interfaces) is the emerging discipline of designing tool interfaces optimised for LLM consumption — as distinct from HCI (Human-Computer Interfaces), which designs for human perception and cognition.
The term is used by Amjad Masad to describe the interface layer Replit built for its coding agent, and referenced as an area of active research with published papers.
Core idea
LLMs are trained on text and reason over text. They are not humans, and their information-processing characteristics differ substantially from human ones:
- They do not “see” a screen; they process token sequences.
- Screenshot-based interfaces (like Anthropic’s computer use) require expensive image processing and are imprecise.
- The optimal granularity and format of feedback differs from what human users need.
- Context window constraints determine how much interface output is useful.
ACI asks: given these characteristics, what is the right interface for an LLM acting as an agent?
Replit’s implementation
Replit built ACI-optimised interfaces for each component of its coding environment:
| Component | Human interface | ACI-optimised equivalent |
|---|---|---|
| Shell | Terminal output (visual, ANSI colours) | Structured text representation at timed increments |
| Editor | Syntax-highlighted file view | Structured diff with inline error signals |
| Package manager | Progress bar, installation logs | Structured install feedback with success/failure flags |
| Services | Dashboard UI | Explicit API endpoints (database, object storage, auth) |
The agent receives text representations designed for efficient model consumption rather than visual representations designed for human comprehension.
Why it matters
The choice of interface materially affects agent performance. Amjad notes this is “more art than science” currently — teams are still experimenting with questions like:
- How many files can you show the model before it starts to hallucinate?
- What granularity of shell output is useful vs. noisy?
- How should errors be surfaced to trigger self-correction?
As agentic coding systems mature, ACI will likely become as systematic as HCI: a set of principles and patterns, not just ad-hoc choices.
Relationship to other concepts
ACI is closely related to Tool Use — the general capability of LLMs to use external tools via structured calls. ACI is the design layer that determines how those tools present information back to the model.
Computer use (Anthropic’s screenshot-based interface) and ACI represent different philosophies: computer use maximises generality (any app, any interface), ACI maximises efficiency (purpose-built for model cognition). Neither is universally superior; the tradeoff is generality vs. precision.
Where mainstream views differ
Some practitioners argue general-purpose interfaces (computer use, browser control) are adequate and will become efficient enough to make purpose-built ACI unnecessary as models improve. Amjad’s position is that purpose-built text interfaces will retain a cost and reliability advantage even as models improve.
See also
- Tool Use
- Agentic Engineering
- Amjad Masad on Replit
- Boris Cherny on Claude Code — related: minimal scaffolding philosophy