agentforce-skills-research

Agentforce Skills Research

Research how to configure, debug, and optimize Agentforce features — topics, actions, instructions, Custom Lightning Types, knowledge grounding, planner behavior, and more. Pulls from external docs, internal repos, Slack, and Google Docs.

Invocation: /agentforce-skills-research [your question]

What This Skill Does

You ask a question about building or configuring Agentforce. The skill:

1. Searches broadly — external docs (Salesforce Help, Trailhead, developer docs), internal repos (git.soma), Slack channels, and Google Docs
2. Finds working examples — real implementations, not just theory
3. Identifies known issues — GUS bugs, platform quirks, workarounds
4. Writes a research artifact — saves findings to .agents/artifacts/<topic>-research.md for future reference
5. Summarizes key findings — concise answer with sources

Example Questions

• "How do I configure action chaining in a topic?"
• "What's the correct way to set up a CLT output rendering?"
• "How does the planner decide which action to run next?"
• "What are the known issues with isConfirmationRequired and chain continuation?"
• "How do I set up knowledge grounding with Data Library?"
• "What's the difference between GenAiFunction and GenAiPlannerBundle metadata?"
• "How do context variables work across actions?"
• "What's the best practice for error handling in Apex actions for Agentforce?"
• "How do I make the agent stop asking 'Can I proceed?' before running an action?"
• "What instruction patterns make the planner more reliable?"

Research Pipeline

Phase 1: Understand the Question

Classify the question into one or more areas:

Phase 2: Search Sources

Search in this priority order. Stop when you have a confident answer — don't search everything if the first source answers it.

Source 1: External Docs (Web Search)

Search: site:help.salesforce.com OR site:developer.salesforce.com OR site:trailhead.salesforce.com
Query: "agentforce" + [specific topic]

Key PDF (download and read locally):
- Agentforce Implementation Guide (Customer-Facing Agents) — 119 pages, updated Apr 2026
  https://resources.docs.salesforce.com/rel1/doc/en-us/static/pdf/agentforce_implementation_customer.pdf
  Local copy: /tmp/agentforce_implementation_customer.pdf
  Covers: Ideate (p.8-20), Build (p.21-60 — subagents, actions, grounding, permissions),
  Test (p.61-91 — planning, core testing, iteration), Deploy (p.92-111 — connections, routing,
  escalation, website), Monitor (p.112-119 — observability, optimization, session tracing)

Good for: Official API docs, metadata reference, setup instructions, release notes, end-to-end implementation guidance.

Source 2: Internal Repos (Codesearch)

Search git.soma for:
- Working Agentforce configurations (genAiPlannerBundles, genAiFunctions)
- Apex action patterns (@InvocableMethod with Agentforce annotations)
- Lightning Type bundles
- Demo repos with topic instructions

Key repos to check:

• chad-goldsmith/pet-travel-demo — full working CLT demo
• chad-goldsmith/sf-demo-skills — demo patterns + best practices
• alberto-ruiz/sra-ga-docs — CX team's SRA GA landing page (all things Service Rep Assistant)

Live: https://git.soma.salesforce.com/pages/alberto-ruiz/sra-ga-docs/

• Search broadly for genAiFunction or GenAiPlannerBundle across all repos

Good for: Working code examples, real configurations, patterns that actually deploy.

Source 3: Slack (Internal Discussions)

Channels to search:
- #agentforce-skills-coworker-collab (C0B2XJ87SBH) — PRIMARY for skills platform updates, APIs, timeline
- #service-assistant-engineering (C06TPK97CCE)
- #sc-service-planner-eng
- #sc-service-planner-leads (C07DVDVH26A)
- #ngs-engineering (C06NDLHQJD7)
- #spa-sf-engineering (C02P450NJ84)
- #help-262-summer26-release
- #soba-engineering (C05UAR03WHY)
- #ai-club (C058L05637W) — cross-cutting AI patterns, agent architecture, tooling, internal adoption

Good for: Known bugs, workarounds, recent changes, "why doesn't X work" answers, platform behavior confirmations, AI architecture patterns.

Source 4: Google Docs (Internal Design Docs)

Search for: "Agentforce" + [topic], "Service Planner" + [topic], "GenAi" + [topic]

Good for: Design decisions, architecture docs, PRDs, internal specifications.

Source 5: Local Knowledge (Already Known)

Check these local files:
- ~/.aisuite/notebook/.agents/artifacts/agentforce-skills-platform-spec.md — OFFICIAL platform spec (Kumar Kasimala)
  → Architecture, SKILL.md frontmatter spec, APIs, data model, shipping OOTB skills, permissions
  → Source doc: https://docs.google.com/document/d/1pOn7zW8P-aYCH0mxKhuOW4FKCgMwE8rmPuX5dZbo29A
- ~/.aisuite/notebook/2026-06-24/agentforce-skills-impact-on-sra.md — SRA impact analysis + architecture proposal findings
  → Source doc: https://docs.google.com/document/d/1n9vr5grKFf1QKkpdtvZeWkIWeHrfFfUdu7oxMkTEgDY
  → Customer-facing agent architecture: live skills, background skills, skill scoping, execute_tool proxy
- ~/.aisuite/notebook/.agents/artifacts/agentforce-skills-customer-agents-architecture.md — Full architecture proposal extraction
  → Execution model, runtime stack, skill scoping, data model, SKILL.md spec, security
- ~/.aisuite/notebook/.agents/artifacts/agentforce-skills-m1-prd.md — M1 Pilot PRD analysis
  → Source doc: https://docs.google.com/document/d/1fE_dz8U1mwPTdyGCk6HwHlGoXPpw0rapt4XNkEanMrc
  → Timeline, scope, 5 global tools, permissions, Skill Builder, customer-facing deferred
- ~/.aisuite/notebook/.agents/artifacts/agentforce-skills-vs-record-companion-actor-framework.md — Skills vs Record Companion comparison
  → Event-driven (RC) vs intent-driven (Skills), convergence points, tension, recommended architecture
  → RC architecture docs: https://docs.google.com/document/d/13hXuGr5o2PZnfgRXE-D85eWNYCOZ1iAWM8sXHsudUlU/edit
- ~/sf-demo-skills/BEST-PRACTICES.md — design patterns and rules
- ~/sf-demo-skills/CLT-GUIDE.md — CLT implementation guide
- ~/sf-demo-skills/SUBAGENT-TEMPLATE.md — subagent creation framework
- ~/pet-travel-demo/skill/REBUILD-AGENT.md — complete action config reference
- ~/pet-travel-demo/skill/TOPIC-INSTRUCTIONS.md — topic instruction example

Key companion specs on git.soma (fetch via codesearch):

• git.soma:kkasimala/skill-builder-specs/SCOPE.md — project scope
• git.soma:kkasimala/skill-builder-specs/specs/day1/10-skill-apis-m1/AGENT_SKILLS_ARCHITECTURE.md — full architecture
• git.soma:kkasimala/skill-builder-specs/specs/day1/10-skill-apis-m1/AGENT_SKILLS_API.md — API reference
• git.soma:kkasimala/skill-builder-specs/Skill_Execution_Agent.yaml — reference agent

Good for: Proven patterns, things already debugged and documented, official platform architecture.

Phase 3: Synthesize & Write

Write findings to: .agents/artifacts/<topic-slug>-research.md

Use this format:

# Research: [Topic Title]
**Date:** [today]
**Query:** [original question]

## Summary
[2-3 sentence answer]

## Key Findings

### [Finding 1 Title]
- **Source:** [where this came from]
- **Detail:** [what was found]
- **Relevance:** [how this answers the question]

### [Finding 2 Title]
...

## Working Examples
[Code snippets, configurations, or screenshots that demonstrate the answer]

## Known Issues / Gotchas
[Platform bugs, non-deterministic behavior, workarounds needed]

## Sources
- [linked source 1]
- [linked source 2]
...

## Still Unknown
[What we couldn't answer — where to look next]

Phase 4: Summarize

Return a concise answer to the user with:

• The key finding (1-2 sentences)
• The most important gotcha or caveat
• Pointer to the full research artifact

Research Depth Levels

Default to Standard unless the user signals otherwise.

Important Rules

• Always cite sources — never present findings without attribution
• Distinguish confirmed from speculative — if a behavior is only observed (not documented), say so
• Date your findings — platform behavior changes across releases. Note the release/date of any source.
• Check the release — Agentforce features change significantly between 262, 264, and beyond. Note which release a finding applies to.
• Don't conflate Guidance Plans and Dynamic Plans — they're different features with different behavior
• "Works in demo" ≠ "works in production" — note when something is demo-only or has reliability caveats
• Save useful findings — even if the user's question is quick, if you find something novel, save it to the artifact folder for future reference