Moss

Appearance

Ampcode Research ​

Agent coding tool by Sourcegraph. Research conducted 2025-12-23.

Core Architecture ​

Minimal agent loop requires only three components:

  1. LLM with tool access
  2. Loop structure (prompt → inference → tool exec → repeat)
  3. Sufficient tokens

Conversation history travels with each request. Server is stateless; client maintains complete context.

Multi-Model Strategy ​

  • Worker (Claude Sonnet 4): Fast, capable, handles bulk of tool use and code generation
  • Oracle (GPT-5 / o3 / Gemini 2.5 Pro): High-level reasoning, architectural planning, debugging
  • Oracle invoked explicitly ("ask the oracle"), not automatically

This is their answer to "which model is best" - use both, specialized by role.

Subagent Architecture ​

Key insight: Subagents multiply context windows.

Instead of spending main agent tokens on a fix, spawn a subagent with fresh 200k context. Only a tiny fraction of main agent tokens used for the handoff.

Properties:

  • Each subagent has isolated context window
  • Full tool access (file edit, terminal, search)
  • Can run in parallel
  • Main agent receives only final summary (no lateral communication)

Model behavior: Claude Sonnet 4 aggressively uses subagents when it spots "clearly defined tasks". Earlier versions (3.5 Sonnet) rarely invoked them.

Specialized subagents:

  • Librarian: Searches remote codebases (all public GitHub + private repos)
  • Generic subagents: Mini-Amps with full capabilities

Context Management ​

"Curated context beats comprehensive context. Every time."

Key decisions:

  • No automatic compaction - risks quality degradation, users notice "agent got dumber"
  • Manual control preferred over automatic threshold-triggered truncation
  • Fixed 200k token window, aggressively leveraged
  • Threads remain immutable; continuation uses explicit "handoff" with new context

AGENTS.md Format ​

Open format for guiding coding agents. 60k+ open-source projects use it.

Structure:

  • Standard Markdown, flexible sections
  • Project overview, setup, build/test commands, code style, testing, security, PR guidelines
  • Closest file in directory tree takes priority
  • Supports YAML frontmatter with glob patterns for conditional inclusion

Precedence: explicit user prompts > AGENTS.md > defaults

Tool Architecture ​

Layers:

  1. Built-in tools: File editing, bash, git
  2. Toolboxes: Scripts responding to TOOLBOX_ACTION env var (describe/execute)
  3. MCP servers: Local/remote, OAuth-capable
  4. Skills: Specialized instruction packages (SKILL.md + resources)

Permission system with sequential rule matching:

  • Actions: allow, reject, ask, delegate
  • Supports regex on tool names and arguments
  • Can delegate to external programs

Edit tool uses string replacement (not diff) - Claude naturally prefers this.

Operational Modes ​

  • Smart: Opus 4.5, unconstrained tokens, 200k context
  • Rush: Faster/cheaper for well-defined tasks
  • Free: OSS models, ad-supported, some limits

Execute mode (-x): Non-interactive, full tool approval for autonomy.

Design Philosophy (from FIF) ​

  • Transparency over abstraction: Show agent work, don't hide complexity
  • Iterative loops over step approval: Review → diagnostics → compiler → tests
  • No VCS manipulation: Agent can't directly git, only shell commands with user direction
  • No .ampignore: File hiding encourages workarounds; use permissions instead
  • No prompt injection defense for Bash (noted as current limitation)

Learnings from Sourcegraph ​

Inversion of control: Provide tools + high-level goals, let agent orchestrate. Don't micromanage via detailed prompts.

Emergent capabilities: Self-correction, test-driven iteration, parallel decomposition emerged from autonomy + feedback loops, not explicit design.

Economics: $5-15 per PR in observed usage. Optimizing for cost suppresses effectiveness.

Human collaboration: Agents implement rough versions, humans refine guardrails. Expertise is "moving guardrails" not prompting.

Background agents: 10-15+ minute tasks. Use CI pipelines as feedback (tests, linters) rather than replicating local env.

Relevance to Moss ​

Potential applications:

  1. Subagent pattern for context multiplication - spawn fresh contexts for isolated tasks
  2. Oracle pattern - dual-model with explicit escalation, not automatic
  3. String replacement over diff for edits (already doing this)
  4. AGENTS.md equivalent - CLAUDE.md serves similar purpose
  5. Curated context - structural tools (view, analyze) align with this philosophy
  6. No compaction - explicit handoffs over automatic truncation

Questions to explore:

  • Should moss TUI support spawning subagents?
  • Oracle pattern worth implementing? (expensive model for stuck situations)
  • Background agent support for long-running analysis?

Sources ​