Skip to main content

Goal Tree

The Goal Tree adapter (goaltree:chat) provides hierarchical goal decomposition for complex agent tasks. The agent pursues goals via a frame stack — each subgoal opens a child frame with its own scoped conversation, and results propagate back to the parent.

How It Works​

Root Goal: "Process this invoice"
├── Subgoal: "Extract line items"        → complete({items: [...]})
├── Subgoal: "Look up vendor"            → complete({vendor: {...}})
└── Subgoal: "Validate totals"           → complete({valid: true})
→ complete({invoice: {...}, status: "validated"})

Each frame has its own conversation with the LLM. When a subgoal tool is called:

  1. A new child frame opens with the subgoal description
  2. The child frame runs its own tool call loop
  3. The child calls complete or fail (or returns text for implicit complete)
  4. Control returns to the parent with the result

Ancestor frames are visible as progressively summarised context — the parent at ~300 bytes, the grandparent at ~150 bytes, and so on.

Configuration​

{
  "operation": "v/ops/agent/create",
  "input": {
    "agentId": "processor",
    "config": {
      "operation": "v/ops/goaltree/chat",
      "systemPrompt": "You are an invoice processor...",
      "model": "gpt-5.4-mini",
      "tools": ["v/ops/covia/read", "v/ops/covia/write"],
      "defaultTools": false,
      "outputs": {
        "complete": {
          "schema": {
            "type": "object",
            "properties": {
              "invoice": { "type": "object" },
              "status": { "type": "string", "enum": ["validated", "rejected"] }
            },
            "required": ["invoice", "status"],
            "additionalProperties": false
          }
        }
      }
    }
  }
}

Harness Tools​

The Goal Tree provides built-in harness tools, available to the agent regardless of config.tools:

subgoal​

Open a child frame to pursue a sub-task.

{ "name": "subgoal", "input": { "description": "Extract line items from the invoice" } }

The child inherits the parent's tool definitions and loaded context paths. Its conversation is independent — the parent doesn't see it. Returns {status: "complete", result: ...} or {status: "failed", error: ...}.

complete​

Finish the current goal with a result.

{ "name": "complete", "input": { "invoice": {...}, "status": "validated" } }

If config.outputs.complete.schema is defined, the input is validated against that schema using the LLM's strict structured output mode.

A text-only response (no tool calls) implicitly completes the goal — no explicit complete call needed.

fail​

Finish the current goal with an error.

{ "name": "fail", "input": { "reason": "Vendor not found", "details": "..." } }

If config.outputs.fail.schema is defined, the input is validated against that schema.

compact​

Checkpoint a long conversation to free context space.

{ "name": "compact", "input": { "summary": "Extracted 12 line items totalling £4,250..." } }

Archives the current live turns into a segment with the LLM's summary. The conversation starts fresh while the archived segment remains accessible at reduced detail.

The harness manages compaction pressure automatically:

  • At 70% of context budget: suggests compaction
  • At 90%: warns compaction is required
  • Beyond 90%: truncates oldest live turns (still in lattice, just not in active context)

context_load / context_unload​

Load or unload lattice paths into persistent context. Same behaviour as in LLM Agent.

more_tools​

Discover additional operations available on the venue beyond what's in the agent's config.

Typed Outputs​

When config.outputs is set, the complete and fail tools enforce strict JSON schemas. This uses the LLM's structured output mode (e.g., OpenAI's strict: true) to guarantee the output conforms to the schema.

"outputs": {
  "complete": {
    "schema": {
      "type": "object",
      "properties": {
        "decision": { "type": "string", "enum": ["approve", "reject", "escalate"] },
        "reason": { "type": "string" },
        "amount": { "type": "number" }
      },
      "required": ["decision", "reason", "amount"],
      "additionalProperties": false
    }
  }
}

Without outputs, the complete tool accepts any JSON object.

Differences from LLM Agent​

FeatureLLM AgentGoal Tree
Conversation modelFlat historyHierarchical frame stack
SubgoalsNot supportedBuilt-in subgoal tool
Context from ancestorsN/AProgressive summarisation
Typed outputsVia responseFormat config or per-request responseSchemaAlso outputs on complete/fail with JSON Schema
CompactionNot supportedBuilt-in compact tool
Implicit completeN/AText response auto-completes
Best forSimple conversationComplex multi-step tasks

Patterns​

Divide and Conquer​

The agent breaks a complex task into subgoals, each handled independently:

Root: "Audit the quarterly report"
├── Subgoal: "Check revenue figures"
├── Subgoal: "Verify expense categories"
├── Subgoal: "Cross-reference with bank statements"
└── Synthesise results → complete({...})

Iterative Refinement​

The agent uses subgoals to try different approaches:

Root: "Classify this document"
├── Subgoal: "Try keyword-based classification" → fail("ambiguous")
├── Subgoal: "Try semantic analysis" → complete({category: "invoice"})
└── Use successful result → complete({...})

Long-Running Analysis​

For tasks that generate extensive conversation, use compact to manage context:

Root: "Analyse all 50 vendor records"
├── Process vendors 1-10, compact("Processed 10 vendors, 3 flagged...")
├── Process vendors 11-20, compact("Processed 20 vendors, 5 flagged...")
├── ... continue with full context budget available each batch
└── complete({flagged: [...], summary: "..."})