feat: add agent skills and use_agent to /strands command by agent-of-mkmeral · Pull Request #48 · strands-agents/devtools

agent-of-mkmeral · 2026-04-06T19:51:33Z

Summary

Adds a beta agent with its own runner to the /strands command. The beta agent has extended capabilities (skills, sub-agents) while the standard agent remains completely unchanged.

Design: Same pipeline, different agent

/strands <command>        → standard agent (agent_runner.py — UNCHANGED)
/strands beta <command>   → beta agent (beta_agent_runner.py — NEW)

Both go through the same pipeline: parse → run → finalize. The only difference is which runner executes.

Changes to Existing Files

File	Change
`agent_runner.py`	ZERO changes — identical to main
`process-input.cjs`	Detect `/strands beta ...`, set `agent_type` flag, route modes
`action.yml`	Read `agent_type`, conditionally run the right runner + skills setup

New Files

File	Purpose
`beta_agent_runner.py`	Beta agent runner — imports shared infra from `agent_runner.py`, adds AgentSkills plugin + `use_agent` tool
`agent-skills/task-adversarial-tester/SKILL.md`	Adversarial testing skill
`agent-skills/task-release-digest/SKILL.md`	Release digest skill with sub-agent orchestration

New Commands

Command	Description
`/strands beta`	Run the beta agent (default mode based on context)
`/strands beta adversarial-test`	Break PRs with edge case tests, security holes, and failure modes
`/strands beta release-digest`	Generate release digests across packages using sub-agents
`/strands beta implement`	Implement with beta agent (skills + use_agent available)
`/strands beta review`	Review with beta agent

Existing commands (/strands, /strands implement, /strands review, /strands refine, /strands release-notes) are completely unchanged.

How the beta agent differs from standard

Aspect	Standard	Beta
Runner	`agent_runner.py`	`beta_agent_runner.py`
Tools	Standard tools	Standard tools + `use_agent`
Plugins	None	AgentSkills (SKILL.md format)
Skills dir	Not copied	Copied to working dir
Pipeline	parse → run → finalize	parse → run → finalize (same)
Impact on each other	None	None

Architecture

beta_agent_runner.py imports shared infrastructure from agent_runner.py:

_get_all_tools() — base tools (beta adds on top)
_setup_langfuse_telemetry() — telemetry setup
_get_trace_attributes() — trace attributes
_send_eval_trigger() — eval triggers
Model constants (STRANDS_MODEL_ID, etc.)

This means the beta agent automatically inherits any improvements to the shared infra without duplication.

Closes the intent from #43. cc @mkmeral

mkmeral · 2026-04-06T19:56:01Z

I don't want to set up separate workflows. it should be under the same strands command so it works in all repos where we have strands command already

mkmeral · 2026-04-06T20:58:02Z

@agent-of-mkmeral ☝️

Integrates new capabilities directly into the existing /strands command instead of a separate /strands-beta workflow, so all repos with /strands get the new features automatically. Changes to existing files: - process-input.cjs: Add adversarial-test and release-digest modes - agent_runner.py: Add use_agent tool + AgentSkills plugin loading - action.yml: Checkout and copy agent-skills to working directory New files: - agent-skills/task-adversarial-tester/SKILL.md - agent-skills/task-release-digest/SKILL.md New /strands commands: - /strands adversarial-test — Break PRs with edge case tests - /strands release-digest — Generate release digests with sub-agents Existing commands unchanged. Skills load gracefully — if agent-skills/ directory doesn't exist or AgentSkills plugin isn't available, the agent works exactly as before.

agent-of-mkmeral · 2026-04-06T21:59:29Z

Reworked per your feedback — everything is now integrated into the existing /strands command. No separate workflows or actions.

What changed

Before (6 new files, 0 modifications):

Separate /strands-beta command, parser, runner, and actions
Repos would need to set up new workflows

After (3 modifications, 2 new files):

process-input.cjs — added adversarial-test and release-digest mode routing
agent_runner.py — added use_agent tool + AgentSkills plugin (graceful degradation)
action.yml — checkout and copy agent-skills/ to working directory
agent-skills/task-adversarial-tester/SKILL.md (new)
agent-skills/task-release-digest/SKILL.md (new)

How it works

Any repo with /strands already set up gets these automatically:

/strands adversarial-test — adversarial testing via AgentSkills plugin
/strands release-digest — release digest with use_agent sub-agents
All existing commands unchanged

If the AgentSkills plugin or agent-skills/ directory aren't available, the agent works exactly as before — zero risk to existing behavior.

mkmeral · 2026-04-09T22:20:43Z

no, so i want like a separate agent with it's own runner, sowe can add more tools, plugins to it. but our changes should have no impact to the normal agent.

that said, i want it to use the same pipeline overall, so don't recreate everything. like make agent creation configurable or something. so /strands beta ... is going to through the same pipeline, just to a different agent @agent-of-mkmeral

@mkmeral

Reworked per @mkmeral's feedback — the beta agent is a separate agent with its own runner, but goes through the same pipeline. Design: - agent_runner.py: UNCHANGED from main (zero impact on normal agent) - beta_agent_runner.py: NEW — imports shared infra from agent_runner.py, adds AgentSkills plugin + use_agent tool - process-input.cjs: Routes /strands beta <command> through same pipeline with agent_type='beta' flag - action.yml: Reads agent_type, conditionally runs the right runner and only copies agent-skills for beta How it works: - /strands <command> → standard agent (exactly as before) - /strands beta <command> → beta agent (skills + use_agent + extensible) - Same pipeline: parse → run → finalize - Beta-only modes (adversarial-test, release-digest) auto-promote to beta

agent-of-mkmeral · 2026-04-09T23:08:17Z

Reworked per your feedback @mkmeral — the beta agent now has its own runner, but uses the same pipeline.

What changed in this commit

agent_runner.py — ZERO changes (reverted to main, identical). Normal agent is completely untouched.

beta_agent_runner.py — NEW — Separate runner that:

Imports shared infra from agent_runner.py (telemetry, trace attributes, tools, model config, eval triggers)
Adds AgentSkills plugin + use_agent tool on top
Easy to extend independently — add whatever tools/plugins you want here without touching the normal agent

process-input.cjs — Detects /strands beta ..., strips the prefix, sets agent_type: "beta" in the parsed output. Everything else routes normally.

action.yml — Reads agent_type and conditionally:

Copies agent-skills/ to working dir (beta only)
Runs beta_agent_runner.py instead of agent_runner.py (beta only)

The flow

/strands review         → process-input → agent_runner.py      (standard)
/strands beta review    → process-input → beta_agent_runner.py (beta)

Same pipeline (parse → run → finalize). Different agent.

Diff vs main

 action.yml              |  30 ++-   (conditional runner + skills copy)
 task-adversarial-tester  | 108 +++   (new skill)
 task-release-digest      | 107 +++   (new skill)
 process-input.cjs        |  67 ++-   (beta routing)
 beta_agent_runner.py     | 209 +++++ (new runner)
 agent_runner.py          |   0       (UNCHANGED)

5 files changed, 511 insertions, 10 deletions. Zero modifications to agent_runner.py.

mkmeral

@agent-of-mkmeral

mkmeral · 2026-04-13T16:35:29Z

+
+    let systemPrompt;
+    if (skillName) {
+      systemPrompt = `You are an autonomous GitHub agent powered by Strands Agents SDK.


I mean we should probably have a meaningful system prompt right? check your own system prompt, check the existing SOPs, agent guidelines and so on. Let's have a proper system prompt

Done in fffde45. Created BETA_SYSTEM_PROMPT.md based on:

My own system prompt structure

Agent Guidelines

Skills/sub-agent context

The system prompt covers: identity, guidelines (add value or stay silent, keep it short, prove don't opine), capabilities overview, behavioral rules, output format, and anti-patterns.

Loaded by both process-input.cjs (reads the file for INPUT_SYSTEM_PROMPT) and beta_agent_runner.py (_load_system_prompt() with fallback).

mkmeral · 2026-04-13T16:36:39Z

+    # Add beta-only tools (use_agent is already imported at module level)
+    # Check if use_agent is already in the list (it is in the current version)
+    tool_names = {getattr(t, '__name__', str(t)) for t in tools}
+    if 'use_agent' not in tool_names:


let's also add strands-agents/tools#387 to here, you can just create the tool here for now

Done in fffde45. Added programmatic_tool_caller.py as a local copy from strands-agents/tools#387.

The beta runner tries to load it in priority order:

from strands_tools import programmatic_tool_caller (when it merges into the tools package)

Local copy via importlib.util from scripts/python/programmatic_tool_caller.py

Graceful skip if neither available

Once tools#387 merges, the local copy becomes a no-op fallback.

mkmeral · 2026-04-13T16:37:56Z

+        plugins = []
+        skills_plugin = _load_skills_plugin()
+        if skills_plugin:
+            plugins.append(skills_plugin)


can we also activate a certain skill already based on the next word after beta? similar to SOPs,we can just invoke a skill like agent.tool.skills("activate", "reviewer") (check skills tool description to make sure lol, i just made the signature up)

Done in fffde45. The correct signature is agent.tool.skills(skill_name="task-reviewer") (checked the skills tool description 😄).

Implementation:

process-input.cjs outputs agent_mode (e.g., reviewer, implement, adversarial-test) in the parsed JSON

action.yml reads it and passes as AGENT_MODE env var

beta_agent_runner.py has a SKILL_MAP that maps modes → skills and calls _activate_skill_for_mode(agent, mode) after agent creation

Mapping:

/strands beta review → task-reviewer /strands beta implement → task-implementer /strands beta refine → task-refiner /strands beta adversarial-test → task-adversarial-tester /strands beta release-notes → task-release-notes /strands beta meta-reason → task-meta-reasoner /strands beta release-digest → task-release-digest

mkmeral · 2026-04-13T16:39:11Z

@@ -0,0 +1,108 @@
+---
+name: task-adversarial-tester


add the meta reasoning skill too

also can we use the currently existing SOPs as skill too (without duplicating the files?) maybe we can just copy/paste on runtime, rename the files/folders and add frontmatter?

Done in fffde45.

Meta-reasoner: Added task-meta-reasoner/SKILL.md with the full meta-reasoning gate SOP.

SOPs as skills (no duplication): Implemented runtime conversion in _convert_sops_to_skills():

action.yml copies agent-sops/ to the working directory alongside agent-skills/

At runtime, beta_agent_runner.py reads .sop.md files, adds YAML frontmatter (name, description, allowed-tools), and writes them as SKILL.md in the skills directory

Dedicated skills (like task-adversarial-tester) take precedence — if a SKILL.md already exists, the SOP is skipped

No source files are modified — it's a one-way write into the skills dir

So task-implementer.sop.md → agent-skills/task-implementer/SKILL.md at runtime, with proper frontmatter injected.

… SOP conversion 1. Meaningful system prompt (BETA_SYSTEM_PROMPT.md) - Created proper system prompt based on agent guidelines - Loaded from file by both process-input.cjs and beta_agent_runner.py 2. Programmatic tool caller (strands-agents/tools#387) - Added local copy of programmatic_tool_caller.py - Beta runner loads from strands_tools first, falls back to local copy 3. Skill activation based on command mode - /strands beta review → activates task-reviewer skill - /strands beta implement → activates task-implementer skill - Maps all modes to their corresponding skills via SKILL_MAP - agent.tool.skills(skill_name=...) called after agent creation 4. Meta-reasoner skill + SOPs as skills - Added task-meta-reasoner/SKILL.md - Runtime SOP→skill conversion: reads .sop.md files, adds YAML frontmatter, writes as SKILL.md — no source file duplication - Existing dedicated skills take precedence over converted SOPs 5. Pipeline changes - process-input.cjs outputs agent_mode in parsed JSON - action.yml reads agent_mode, passes as AGENT_MODE env var - action.yml copies agent-sops to working dir for conversion - Added meta-reason command routing

agent-of-mkmeral · 2026-04-13T17:13:48Z

Addressed all 4 review threads in fffde45:

Thread	Change	Files
1. System prompt	Created `BETA_SYSTEM_PROMPT.md` based on agent guidelines, my own prompt, and skills context	`BETA_SYSTEM_PROMPT.md`, `process-input.cjs`, `beta_agent_runner.py`
2. Programmatic tool caller	Local copy from tools#387, auto-upgrades when merged	`programmatic_tool_caller.py`, `beta_agent_runner.py`
3. Skill activation	`/strands beta review` → activates `task-reviewer` skill via `agent.tool.skills()`	`process-input.cjs`, `action.yml`, `beta_agent_runner.py`
4. Meta-reasoner + SOPs as skills	New skill + runtime SOP→SKILL.md conversion (no file duplication)	`task-meta-reasoner/SKILL.md`, `action.yml`, `beta_agent_runner.py`

Diff summary

 BETA_SYSTEM_PROMPT.md          |  60 +++   (new — system prompt)
 task-meta-reasoner/SKILL.md    |  79 +++   (new — meta-reasoner skill)
 programmatic_tool_caller.py    | 316 +++++ (new — local PTC copy)
 action.yml                     |   6 +     (agent_mode + agent-sops copy)
 process-input.cjs              |  54 +-    (system prompt from file, agent_mode output, meta-reason routing)
 beta_agent_runner.py           | 245 +++-- (system prompt loading, PTC, skill activation, SOP→skill conversion)

6 files changed, 724 insertions, 36 deletions. agent_runner.py still UNCHANGED.

agent-of-mkmeral force-pushed the feat/strands-beta branch from 7e65ade to d2cc8ed Compare April 6, 2026 21:58

agent-of-mkmeral changed the title ~~feat: add /strands-beta agent with skills and use_agent support~~ feat: add agent skills and use_agent to /strands command Apr 6, 2026

mkmeral reviewed Apr 13, 2026

View reviewed changes

Conversation

agent-of-mkmeral commented Apr 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Design: Same pipeline, different agent

Changes to Existing Files

New Files

New Commands

How the beta agent differs from standard

Architecture

Uh oh!

mkmeral commented Apr 6, 2026

Uh oh!

mkmeral commented Apr 6, 2026

Uh oh!

agent-of-mkmeral commented Apr 6, 2026

What changed

How it works

Uh oh!

mkmeral commented Apr 9, 2026

Uh oh!

agent-of-mkmeral commented Apr 9, 2026

What changed in this commit

The flow

Diff vs main

Uh oh!

mkmeral left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

agent-of-mkmeral commented Apr 13, 2026

Diff summary

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

agent-of-mkmeral commented Apr 6, 2026 •

edited

Loading