AgentLang

Uses markdown heading hierarchy as structural skeleton and numbered lists as execution steps. Friendly to non-programmers, unambiguous to agents.

---

Keywords

8 keywords covering all control flow:

| Keyword | Purpose | Example | |---------|---------|---------| | def: | Define an action | #### def: FetchData(url) | | call | Invoke an action (local or other agent) | call FetchData: get user list | | if | Conditional | if fails: return | | else | Alternative branch | else: retry | | repeat | Loop | repeat until done | | parallel | Parallel execution | parallel: + sub-step list | | confirm | Pause for user confirmation or input | confirm: continue? | | return | Terminate current action and output result | return data |

---

Writing Rules

1. ### [Entrypoint: agentlang] — H3 + brackets = application entry, identifies this as an AgentLang program
2. #### def: Name — H4 + def: = action definition
3. #### Main — H4 + Main = execution entry
4. Numbered lists = sequential steps
5. parallel: followed by - unordered list = parallel steps
6. Description after colon = optional intent context for the agent

---

Parsing Guide

Parsing rules (execute in order):

1. Scan all #### def: to build an action index
2. Start sequential execution from #### Main
3. On call X → resolve as "execute X", no distinction between local def and external agent (unified model)
4. On call X: description or call X("arg"): description → description serves as execution context
5. On parallel: → execute child - list items concurrently
6. On confirm → pause execution, wait for user response; if user declines, terminate current flow
7. Natural language in steps → agent interprets and executes autonomously
8. Parameters: inferred from context by default; explicit passing uses Name(arg0, arg1) form

Structure Recognition

| Markdown Element | Semantics | |-----------------|-----------| | ### [Entrypoint: agentlang] | Application entry, identifies an AgentLang program | | #### def: Name | Action definition start | | #### def: Name(args) | Action definition with parameters | | #### Main | Execution entry | | 1. 2. 3. numbered list | Sequential steps | | - unordered list (under parallel:) | Parallel steps |

Execution Model

• Steps execute in numbered order unless parallel: is encountered
• call is the sole invocation mechanism, handling both local actions and external agents uniformly
• confirm pauses execution flow, waits for user response; user decline terminates current flow
• if / else provide branching, agent decides based on runtime state
• repeat creates loops, requires termination condition (until, attempt limit)
• return terminates current action, outputs result to caller
• Each step's return value is implicitly passed to the next step (unless explicitly specified)

---

Example

[Entrypoint: agentlang]

def: ParseSkill(source)

1. scan source for ### [Entrypoint: agentlang] to identify skill boundary
2. scan all #### def: headings to build action index
3. locate #### Main as execution entry
4. return parsed skill structure

def: ExecuteStep(step)

1. if step starts with call: resolve target action and execute it
2. if step starts with confirm: pause and wait for user response; if declined, terminate flow
3. if step starts with parallel:: execute child list items concurrently
4. if step starts with if: evaluate condition and branch
5. if step starts with repeat: loop until termination condition met
6. else: interpret natural language and execute

Main

1. call ParseSkill: parse the given AgentLang source document
2. read the first step from Main block
3. repeat until all steps are processed:
   • call ExecuteStep: execute current step
   • if step returns error: return error
   • advance to next step
4. return final result