Forage Solutions

Explore solution space via ant colony opt — deploy indep hypotheses as scouts, reinforce promising via evidence, detect diminishing returns, know when abandon strategy + explore elsewhere.

Use When

• Problem w/ multiple plausible approaches, no clear winner
• First approach not working but alts unclear
• Debug w/ no obvious root cause — multiple hypotheses need parallel
• Search codebase for behavior source, location unknown
• Previous attempts converged prematurely suboptimal
• Complement build-coherence when space must be explored before decision

In

• Required: Problem/goal (what foraging for?)
• Required: Current knowledge state (what already known?)
• Optional: Previous approaches + outcomes
• Optional: Exploration constraints (time, tool availability)
• Optional: Urgency (affects explore-exploit balance)

Do

Step 1: Map Landscape

Before deploying, characterize shape.

Solution Distribution Types:
┌────────────────────┬──────────────────────────────────────────────────┐
│ Type               │ Characteristics and Strategy                     │
├────────────────────┼──────────────────────────────────────────────────┤
│ Concentrated       │ One correct answer exists (bug fix, syntax       │
│ (one right fix)    │ error). Deploy many scouts quickly to locate     │
│                    │ it. Exploit immediately when found               │
├────────────────────┼──────────────────────────────────────────────────┤
│ Distributed        │ Multiple valid approaches (architecture choice,  │
│ (many valid paths) │ implementation strategy). Scouts assess quality  │
│                    │ of each. Use `build-coherence` to choose         │
├────────────────────┼──────────────────────────────────────────────────┤
│ Ephemeral          │ Solutions depend on timing or sequence (race     │
│ (time-sensitive)   │ conditions, order-dependent bugs). Fast scouting │
│                    │ with immediate exploitation. Cannot revisit       │
├────────────────────┼──────────────────────────────────────────────────┤
│ Nested             │ Solving the surface problem reveals a deeper one │
│ (layers of cause)  │ (config issue masking an architecture problem).  │
│                    │ Scout at each layer before committing to depth   │
└────────────────────┴──────────────────────────────────────────────────┘

Classify. Distribution type → how many scouts + how fast switch exploration → exploitation.

→ Clear characterization informs scouting. Feels accurate not forced.

If err: completely unknown → itself = classification. Treat as potentially distributed + deploy broad scouts. First round reveals character.

Step 2: Deploy Scout Hypotheses

Gen indep hypotheses as scouts. Each probes diff direction.

1. Gen 3-5 indep hypotheses about problem/solution
2. Each → 1 cheap test (single file read, 1 grep, 1 check)
3. Rate initial promise on evidence (not gut)
4. Deploy indep: no let A influence test of B

Scout Deployment Template:
┌───────┬──────────────────────┬──────────────────────┬──────────┐
│ Scout │ Hypothesis           │ Test (one action)    │ Promise  │
├───────┼──────────────────────┼──────────────────────┼──────────┤
│ 1     │                      │                      │ High/Med/│
│ 2     │                      │                      │ Low      │
│ 3     │                      │                      │          │
│ 4     │                      │                      │          │
│ 5     │                      │                      │          │
└───────┴──────────────────────┴──────────────────────┴──────────┘

Key: scouts assess not exploit. Quick signal each, not deep investigation first promising.

→ 3-5 indep hypotheses + cheap tests. None deeply explored yet — breadth-first pass.

If err: <3 hypotheses → (a) very constrained (concentrated — good, scout aggressive) or (b) understanding too shallow (read more context). Hypotheses not indep (variations same) → too narrow, force ≥1 contradicting others.

Step 3: Trail Reinforcement — Follow Evidence

After scout results, reinforce promising, let weak decay.

1. Review results: which found supporting evidence?
2. Strong evidence → reinforce: invest more investigation
3. No evidence → decay: don't investigate w/o new signals
4. Contradicting → inhibition: actively avoid
5. Monitor premature convergence: all effort to first reinforced → force 1 scout into unexplored

Trail Reinforcement Decision:
┌───────────────────────────┬──────────────────────────────────────┐
│ Scout Result              │ Action                               │
├───────────────────────────┼──────────────────────────────────────┤
│ Strong supporting evidence│ REINFORCE — deepen investigation     │
│ Weak supporting evidence  │ HOLD — one more cheap test before    │
│                           │ committing                           │
│ No evidence               │ DECAY — deprioritize, scout elsewhere│
│ Contradicting evidence    │ INHIBIT — mark as dead end           │
│ Ambiguous result          │ REFINE — hypothesis was too vague,   │
│                           │ sharpen and re-scout                 │
└───────────────────────────┴──────────────────────────────────────┘

→ Clear prioritization on evidence not preference. Strongest gets most but ≥1 alt alive.

If err: all empty → hypotheses wrong, not approach. Reframe: "What assumptions could be wrong?" Gen new from diff angle. All strong → distributed (multiple valid) → build-coherence for selection.

Step 4: Marginal Value Theorem — Know When Leave

Monitor yield. Info per effort drops below avg across all → switch.

Marginal Value Assessment:
┌────────────────────────┬──────────────────────────────────────────┐
│ Signal                 │ Action                                   │
├────────────────────────┼──────────────────────────────────────────┤
│ New information per    │ CONTINUE — this trail is productive      │
│ action is high         │                                          │
├────────────────────────┼──────────────────────────────────────────┤
│ New information per    │ PREPARE TO SWITCH — squeeze remaining    │
│ action is declining    │ value, begin scouting alternatives       │
├────────────────────────┼──────────────────────────────────────────┤
│ Last 2-3 actions       │ SWITCH — the trail is depleted. The cost │
│ yielded nothing new    │ of staying exceeds the cost of switching │
├────────────────────────┼──────────────────────────────────────────┤
│ Information contradicts│ SWITCH IMMEDIATELY — not just depleted   │
│ earlier findings       │ but misleading. Cut losses               │
└────────────────────────┴──────────────────────────────────────────┘

Important: factor switching cost. Moving to new hypothesis = loading new context = cost. Don't switch marginal gains → only when clearly depleted.

→ Deliberate continue or switch on yield assessment, not habit/frustration. Switches evidence-based not impulse.

If err: switching too frequent (oscillation) → switching cost undervalued. Commit to current N more actions before reassess. Never switching (stuck despite declining) → hard cap: after N unproductive, switch regardless sunk cost.

Step 5: Adapt Strategy

Based on results → select next phase.

1. Most empty, one weak → misframed. Step back + reframe: what question?
2. One strong, others empty → concentrated. Exploit strong w/ full attention
3. Multiple competing → distributed. build-coherence to select
4. Clear winner emerging → explore → exploit. Reduce scouting budget 10-20% (keep 1 scout active alts), commit primary effort to winning
5. All exhausted → solution may not exist in current space. Expand: diff tools, diff assumptions, ask user

→ Strategic decision next phase follows logically from results. Feels like conclusion not guess.

If err: no strategy feels right → foraging revealed genuine uncertainty, valid outcome. Communicate to user: "Explored N, found X. Most promising Y because Z. Pursue or additional context?"

Check

• Landscape characterized before scouting
• ≥3 indep hypotheses gen + tested
• Tests cheap (1 action each) + indep
• Reinforcement on evidence not preference
• Marginal value assessed before deep investigation
• Strategy adapted to results not fixed plan

Traps

• Premature exploitation: Dive deep first showing any promise w/o scouting alts. Most common — first good idea often not best.
• Perpetual scouting: Gen hypotheses endless never commit. Set budget: after N scouts, commit best regardless.
• Non-indep hypotheses: "Maybe in file A" + "maybe in file B imported by A" = not indep, share assumptions. Force genuine diversity.
• Ignore inhibition: Evidence contradicts → let go. Continue investing contradicted because effort spent = sunk cost fallacy.
• Scout w/o record: Not recorded → later scouts repeat. Briefly note each scout finding before moving.

→

• forage-resources — multi-agent foraging model this adapts to single-agent
• build-coherence — foraging reveals multiple valid needing eval
• coordinate-reasoning — manages info flow between scout hypotheses + exploitation
• awareness — monitors premature convergence + tunnel vision during foraging