docs: agent design principles — determinism/judgment split (#240)

Design principle for all disinto agents.

## Core idea
Split every agent into two layers:
- **Bash orchestrator (thin, deterministic):** session lifecycle, worktrees, locks, phase monitoring
- **Claude via formula (fat, judgment):** understand task, implement, handle reviews/CI/merge, adapt to novel situations

## Why
Agent scripts grow by accretion — every lesson becomes another if/else in bash. Formulas are refineable, learnable, and generalizable. Bash state machines are not.

## Includes
- Design principle with clear split criteria
- "When reviewing, ask these questions" checklist
- Current state assessment for all 5 agent types
- Risk mitigations (phase protocol as safety net)

Reviewers and planner should be aware of this principle when assessing PRs and planning work.

Co-authored-by: openhands <openhands@all-hands.dev>
Reviewed-on: https://codeberg.org/johba/disinto/pulls/240
Reviewed-by: Disinto_bot <disinto_bot@noreply.codeberg.org>
This commit is contained in:
johba 2026-03-19 09:56:37 +01:00
parent a4877b58db
commit b9ba5c9250

110
docs/AGENT-DESIGN.md Normal file
View file

@ -0,0 +1,110 @@
# Agent Design Principles
> **Status:** Active design principle. All agents, reviewers, and planners should follow this.
## The Determinism / Judgment Split
Every agent has two kinds of work. The architecture should separate them cleanly.
### Deterministic (bash orchestrator)
Mechanical operations that always work the same way. These belong in bash scripts:
- Create and destroy tmux sessions
- Create and destroy git worktrees
- Phase file watching (the event loop)
- Lock files and concurrency guards
- Environment setup and teardown
- Session lifecycle (start, monitor, kill)
**Properties:** No judgment required. Never fails differently based on interpretation. Easy to test. Hard to break.
### Judgment (Claude via formula)
Operations that require understanding context, making decisions, or adapting to novel situations. These belong in the formula — the prompt Claude executes inside the tmux session:
- Read and understand the task (fetch issue body + comments, parse intent)
- Assess dependencies ("does the code this depends on actually exist?")
- Implement the solution
- Create PR with meaningful title and description
- Read review feedback, decide what to address vs push back on
- Handle CI failures (read logs, decide: fix, retry, or escalate)
- Choose rebase strategy (rebase, merge, or start over)
- Decide when to refuse vs implement
**Properties:** Benefits from context. Improves when the formula is refined. Adapts to novel situations without new bash code.
## Why This Matters
### Today's problem
Agent scripts grow by accretion. Every new lesson becomes another `if/elif/else` in bash:
- "CI failed with this pattern → retry with this flag"
- "Review comment mentions X → rebase before addressing"
- "Merge conflict in this file → apply this strategy"
This makes agents brittle, hard to modify, and impossible to generalize across projects.
### The alternative
A thin bash orchestrator handles session lifecycle. Everything that requires judgment lives in the formula — a structured prompt that Claude interprets. Learnings become formula refinements, not bash patches.
```
┌─────────────────────────────────────────┐
│ Bash orchestrator (thin, deterministic) │
│ │
│ - tmux session lifecycle │
│ - worktree create/destroy │
│ - phase file monitoring │
│ - lock files │
│ - environment setup │
└────────────────┬────────────────────────┘
│ inject formula
┌─────────────────────────────────────────┐
│ Claude in tmux (fat formula, judgment) │
│ │
│ - fetch issue + comments │
│ - understand task │
│ - assess dependencies │
│ - implement │
│ - create PR │
│ - handle review feedback │
│ - handle CI failures │
│ - rebase, merge, or escalate │
└─────────────────────────────────────────┘
```
### Benefits
- **Adaptive:** Formula refinements propagate instantly. No bash deploy needed.
- **Learnable:** When an agent handles a new situation well, capture it in the formula.
- **Debuggable:** Formula steps are human-readable. Bash state machines are not.
- **Generalizable:** Same orchestrator, different formulas for different agents.
### Risks and mitigations
- **Fragility:** Claude might misinterpret a formula step → Phase protocol is the safety net. No phase signal = stall detected = supervisor escalates.
- **Cost:** More Claude turns = more tokens → Offset by eliminating bash dead-ends that waste whole sessions.
- **Non-determinism:** Same formula might produce different results → Success criteria in each step make pass/fail unambiguous.
## Applying This Principle
When reviewing PRs or designing new agents, ask:
1. **Does this bash code make a judgment call?** → Move it to the formula.
2. **Does this formula step do something mechanical?** → Move it to the orchestrator.
3. **Is a new `if/else` being added to handle an edge case?** → That's a formula learning, not an orchestrator feature.
4. **Can this agent's bash be reused by another agent type?** → Good sign — the orchestrator is properly thin.
## Current State
| Agent | Lines | Judgment in bash | Target |
|-------|-------|------------------|--------|
| dev-agent | 1380 (agent 732 + phase-handler 648) | Heavy — deps, CI retry, review parsing, merge strategy, recovery mode | Thin orchestrator + formula |
| review-agent | 870 | Heavy — diff analysis, review decision, approve/request-changes logic | Needs assessment |
| supervisor | 877 | Heavy — multi-project health checks, CI stall detection, container monitoring | Partially justified (monitoring is deterministic, but escalation decisions are judgment) |
| gardener | 1242 (agent 471 + poll 771) | Medium — backlog triage, duplicate detection, tech-debt scoring | Poll is heavy orchestration; agent is prompt-driven |
| vault | 442 (4 scripts) | Medium — approval flow, human gate decisions | Intentionally bash-heavy (security gate should be deterministic) |
| planner | 382 | Medium — AGENTS.md update, gap analysis | Migrating to tmux+formula (#232) |
| action-agent | 192 | Light — formula execution | Close to target |