Guide

What Open Horizons Adds

Open Horizons keeps an agent run from becoming “ask for patch, accept patch.”

Source: docs/open-horizons.md

What Open Horizons Adds

Open Horizons keeps an agent run from becoming “ask for patch, accept patch.”

For LLM development, the sequence has four arcs:

ArcSteps
GroundingIntent Engineering → model-fit note → context pack
Framing/aim → /problem-space → /problem-statement → /solution-space
Delegationevidence → brief → skill decision → subagent decision → /execute
Reflection/review → /dissent → knowledge extraction → /salvage

The reason is simple: do not assign work until you know what good means, what language operation the model should perform, what context it should inherit, what role boundary it needs, and what should survive after the session.

Evidence first

Here, evidence is not the agent's confidence.

Evidence is:

  • failing behavior reproduced;
  • tests that fail before the fix and pass after;
  • static checks;
  • build output;
  • review findings;
  • deleted duplicate paths;
  • clarified ownership;
  • a smaller blast radius;
  • a salvage note that prevents repeating the same mistake.

What each tool does

ToolJob in this repo
Intent EngineeringClarify intent, burst, pause, structure, and iterate.
Model-fit noteName the model job — extract, compare, classify, rewrite, critique, generate candidates, or translate — and the context it must be supplied rather than asked to guess.
Context packPreserve selective context before delegation.
/aimName the outcome before the agent ranks work.
/problem-spaceMap terrain: systems, stakeholders, constraints, blast radius, assumptions.
/problem-statementChoose the framing that points to the right class of fix.
/solution-spaceCompare Band-Aid, Local Optimum, Reframe, and Redesign paths.
Evidence checklistDefine checks before implementation.
Project skillPreserve a repeated procedure as SKILL.md.
SubagentPreserve a role boundary with scoped tools and isolated context.
/executeImplement one selected slice from a brief.
/reviewJudge the patch against the aim and checks.
/dissentLook for the way the accepted-looking patch still fails.
Knowledge extractionRecord metis, signals, guardrails, outcome updates, or ADRs.
/salvageKeep learning when the run drifted and restart smaller.

Strategy for one slice

The agent brief should carry four fields:

FieldQuestion
AimWhat outcome are we trying to create?
MechanismWhy should this approach move that outcome?
FeedbackWhat signal tells us quickly if it is wrong?
GuardrailWhat must not break while we move?

If the brief only says what to change, it is missing the strategy.

Solution levels

This is where Open Horizons meets Beyond the Nearest Peak.

A real project problem usually has more than one altitude:

  • Band-Aid: suppress the symptom;
  • Local Optimum: improve the current design;
  • Reframe: change the problem statement;
  • Redesign: make the class of failure harder to create.

The selected level should show up in the brief and in review.

If you choose a Band-Aid, say why speed or risk makes that acceptable. If you choose Redesign, say why the recurrence justifies the blast radius.

Signals at different scales

ScaleSignal
CodeRegression tests, integration tests, static checks, deleted duplicate paths.
Agent runDid it read the right files, follow the brief, run checks, and stop on drift?
SkillDid the repeated procedure become easier to invoke correctly?
SubagentDid the role boundary produce better evidence or lower context load?
Knowledge artifactDid a future run inherit a verified learning instead of rediscovering it?
ReviewDid the patch move the aim at the selected solution level?
PracticeWhich rejected path, check, guardrail, or metis should survive into the next run?

Shrink a step if needed. Do not remove the signal.

Exercise

Take a planned agent run and mark where each gate happens:

  1. intent;
  2. model-fit note;
  3. context pack;
  4. aim;
  5. problem space;
  6. problem statement;
  7. solution search;
  8. evidence;
  9. brief;
  10. skill decision;
  11. subagent decision;
  12. execute;
  13. review;
  14. dissent;
  15. extraction;
  16. salvage.

If a gate has no artifact, decide whether it is unnecessary for this slice or whether the run is relying on implicit judgment.

Go deeper

← Prompt and Context AssemblyProblem Space →