Agentic performance source library

How to use this library

This is not a list of “best agent platforms.” It is a source map for the seven performance levers in Skills in the agentic performance system:

1. Purpose — what good looks like;
2. Intelligence — which model configuration can do the work;
3. Guidance — what method it should follow;
4. Grounding — what it should know now;
5. Action — what it can see, use, or change;
6. Execution — how the work runs;
7. Assurance — how people know, diagnose, and improve.

Security, privacy, identity, governance, accessibility, and human authority form a trust boundary around all seven.

The sources are deliberately labeled by evidence type:

For presentation citations, prefer the first four categories for factual claims. Use O'Reilly to make the ideas approachable and to surface tensions worth discussing.

A twelve-resource starting shelf

For an audience that should not have to absorb the full bibliography, begin here:

1. Building effective agents — Anthropic for the difference between fixed workflows and agents, and for the “augmented LLM” mental model.
2. A practical guide to building agents — OpenAI for a beginner-friendly progression through models, tools, instructions, orchestration, and guardrails.
3. Agentic AI patterns and workflows — AWS Prescriptive Guidance for reusable architecture patterns and the perception–reason–action loop.
4. Agentic AI architecture guides — Google Cloud for a browsable map from single-agent systems to multi-agent, enterprise data, and security scenarios.
5. AI agent orchestration patterns — Microsoft for explicit tradeoffs among sequential, concurrent, group-chat, handoff, and manager-led patterns.
6. Agentic AI Foundation — current projects and working groups for the vendor-neutral open infrastructure landscape.
7. Agent Skills specification for the portable structure of procedural guidance.
8. Model Context Protocol architecture for the distinction between hosts, clients, servers, tools, resources, prompts, consent, and security boundaries.
9. Effective context engineering for AI agents — Anthropic for treating context as a finite system resource rather than an ever-growing prompt.
10. Demystifying evals for AI agents — Anthropic for tasks, trials, graders, traces, outcome state, and the model–harness pairing.
11. Evaluate agent workflows — OpenAI for trace-based evaluation and regression workflows.
12. What We Learned from a Year of Building with LLMs, Part I — O'Reilly Radar for an accessible practitioner bridge from prompting to retrieval, workflows, evaluation, and monitoring.

Agentic AI Foundation and the Linux Foundation

The Agentic AI Foundation (AAIF) is useful here for a specific reason: it provides a neutral home for open projects and working groups at the interfaces between agentic systems. It should not be presented as a complete reference architecture, certification authority, or guarantee of quality.

The Linux Foundation launch announcement identifies MCP, goose, and AGENTS.md as the three founding project contributions in December 2025. The current AAIF organization page now also lists agentgateway and Agent2Agent (A2A). That distinction matters: “founding projects” and “current hosted projects” are not the same claim.

Projects and open infrastructure

MCP, A2A, and Skills solve different problems. MCP standardizes access to context and capabilities. A2A standardizes inter-agent communication. Skills package procedural guidance. None of them, by itself, supplies clean data, correct permissions, a good professional method, or evidence that the workflow succeeds.

Working groups worth watching

AAIF's working groups are especially relevant because their scopes closely match the outer trust and assurance layers of the proposed visual:

These repositories are active community work, not finished standards. Cite them as places where the relevant problems are being defined and coordinated.

Ecosystem research and discovery

Amazon Web Services

AWS has unusually broad material spanning design patterns, security architecture, runtime services, evaluation, and production operations. The general architecture guidance is portable; the AgentCore implementation details are AWS-specific.

Anthropic

Anthropic's engineering writing is strongest when used as a set of design and evaluation lessons, not as universal laws. Several articles make the model–harness–environment coupling unusually explicit.

OpenAI

OpenAI's current developer documentation is most useful when read as a connected set: model selection, orchestration, Skills, context management, traces, datasets, and graders are parts of one optimization loop.

Google

Google's sources are valuable at three levels: open framework and protocol documentation, managed runtime architecture, and controlled research on multi-agent scaling and memory.

Microsoft

Microsoft's architecture material is helpful because it repeatedly distinguishes model behavior from workload design and treats nondeterminism as an operational concern.

Research papers and benchmarks

A note about arXiv

arXiv is an indispensable public repository, but appearance on arXiv does not mean a paper has been peer reviewed. The table therefore identifies publication status only where the paper's own record makes it clear. Even peer-reviewed benchmarks should be treated as diagnostic instruments rather than universal rankings: agent performance depends on the task distribution, tools, harness, runtime, budgets, and graders.

The last two papers are unusually relevant to this framework because they test its central claim: the useful comparison unit is often the configured agent system, not the model name alone. They are also recent preprints, so they should inspire evaluation design rather than settle the question.

O'Reilly Radar

O'Reilly Radar is most valuable here as a bridge between technical evidence and an audience learning how to think about agentic systems. Use these articles for narrative, questions, and practitioner heuristics; use the primary sources above for specifications and strong factual claims.

Durable foundations

Newer agent-specific perspectives

Independent trust and lifecycle baselines

Vendor documentation explains how to build on a platform; independent frameworks help keep the conversation anchored in organizational responsibility.

What the sources collectively support

Across foundations, cloud providers, model providers, research papers, and practitioner writing, several themes recur strongly enough to anchor the presentation:

1. The model is a component, not the outcome. The harness, instructions, context, tools, environment, and graders change what the model can accomplish.
2. Start with the work, not the architecture. Define the user, end state, acceptable variation, authority, and review path before selecting an agent pattern.
3. Prefer the least complicated structure that passes the evals. A deterministic function, fixed workflow, or single agent is often better than a multi-agent system when the task does not benefit from added autonomy or parallelism.
4. Context is selected, not merely accumulated. Better performance depends on relevance, authority, freshness, placement, compaction, and memory policy—not just context-window size.
5. Tools and software interfaces are part of intelligence in practice. A model cannot reliably overcome ambiguous schemas, overlapping tool descriptions, weak errors, unsafe permissions, or an unstable interface.
6. Protocols improve interoperability, not correctness. MCP, A2A, and similar standards reduce bespoke integration, while data quality, method quality, authorization, and evaluation remain separate responsibilities.
7. Evaluate outcomes and trajectories. End-state checks answer whether the job was completed; traces reveal where selection, retrieval, reasoning, tool use, state, or recovery failed.
8. Variation is a system property. Multiple trials, representative environments, pinned configurations, and runtime controls are necessary for meaningful comparisons.
9. Trust belongs around the system. Identity, least privilege, confirmation, audit, reversibility, privacy, and human authority shape every lever rather than appearing as a final gate.
10. Cost should be optimized against a passing quality bar. Tokens, latency, retries, infrastructure, and human review are valuable metrics only in relation to a correct and useful outcome.

Reading paths for different audiences

New to agentic workflows — approximately 60 minutes

1. Building effective agents
2. A practical guide to building agents
3. AAIF current projects
4. AI agent orchestration patterns
5. The first half of Demystifying evals for AI agents

The inspirational message is: you do not need the entire stack to begin; you need one valuable outcome and enough structure to learn whether it worked.

Designing Skills and reusable guidance

1. Agent Skills specification
2. OpenAI Skills documentation
3. AGENTS.md
4. Effective context engineering
5. Writing effective tools

Architecture and production operations

1. AWS agentic patterns
2. Google agentic architecture guides
3. Microsoft AI workload documentation
4. Anthropic long-running harness guidance
5. OpenAI orchestration

Evaluation and consistency

1. Anthropic agent evals
2. OpenAI agent evals
3. Amazon's evaluation lessons
4. Google Agents CLI evaluation guide
5. $\tau$-bench, SWE-bench, and Harness-Bench

Citation hygiene for the final guide and presentation

• Link to the exact page that supports the claim, not a search result or company home page.
• Add the publisher and, for time-sensitive product documentation, the access or verification date.
• Distinguish a founding project from a project currently hosted by the same foundation.
• Do not call an arXiv preprint “peer reviewed” unless a venue is explicitly confirmed.
• Describe a landscape as a discovery tool, not an endorsement list.
• Treat vendor benchmark findings as evidence about the tested configuration, not every model or workflow.
• Prefer measured statements such as “in this evaluation” over universal statements such as “multi-agent systems are better.”
• Recheck current product, protocol, and working-group pages before publication; this ecosystem changes quickly.