Agentic AI · 2026
AI has shifted from a passive, prompt-response framework into an autonomous, executing force known as agentic AI. Instead of only generating text or answering questions, agentic systems use large language models as a "central nervous system" to independently plan, use tools, collaborate with other agents, and run multi-step workflows.
Pillar 1 of 4
Where agents came from, the standards that define a good one in 2026, and the tools and docs that implement those standards.
1.1 · Evolution
Four eras, from rigid rules to models that plan and act. Under each one: where our repo's work sits.
Loading the timeline…
Powered by rigid if-then structures, basic scripts and early Robotic Process Automation (RPA). These systems could only handle highly predictable tasks and had no real decision-making capability.
Driven by intent-recognition machine learning, which gave rise to voice assistants like Siri and basic customer-service chatbots. They were context-aware but limited to answering single questions without real-world utility.
LLMs like GPT-3 and GPT-4 let machines generate new content and code and reason through logic puzzles. But it stayed System 1 thinking: reactive, prompt-dependent, with no long-term memory and no real-world read-write access.
LLMs are packaged inside cognitive loops. They take an open-ended goal, break it into a task list (planning), fetch documents (RAG), write code, call third-party APIs (tool use), review their own errors, and adjust until the goal is reached.
Generative AI
Prompt in, text out. Reactive, no memory, no access to the world.
Agentic AI
Goal in, outcome out: plan → act with tools → observe → review → repeat.
1.2 · Standards
What an agentic system is expected to do in 2026. Next to each standard is what our repo does about it, pulled from the tasks in our build phases.
Loading…
An open, plug-and-play architecture that lets LLMs safely read data from local files, connect to databases, and invoke software environments in a uniform way. MCP connects an agent to its tools and data.
The standard for agents talking to other agents. Each agent publishes an Agent Card that describes its skills; other agents find it and hand it tasks over HTTP and JSON-RPC, with streamed updates for long jobs, whatever framework each side is built on. MCP connects an agent to tools; A2A connects agents to each other.
ACP (Agent Communication Protocol), started by IBM's BeeAI team, merged into A2A under the Linux Foundation in August 2025. Its repo is archived and IBM published a migration guide.
AGP (Agent Gateway Protocol) came from Cisco's AGNTCY project, not IBM. It has been renamed SLIM, a secure messaging layer that can carry A2A and MCP traffic. AGNTCY is now a Linux Foundation project too.
Watch the name clash: "ACP" today often means Zed's Agent Client Protocol, which connects coding agents to code editors.
Predictable schema formats, like JSON, so agents don't hallucinate arguments when they trigger code execution.
Instead of one monolithic agent doing everything, work flows through standard roles: a Planner splits tasks, a Researcher calls APIs, an Executor runs code, and a Reviewer checks output quality.
Agents are no longer stateless. Short-term working memory lives in scratchpads; long-term memory lives in vector databases that track preferences and history.
Agents verify their software identity with OAuth and dedicated access tokens before they query databases.
Agents run in isolated code sandboxes with least-privilege permissions, for example a cap on how many tools an agent can invoke before it has to re-authenticate.
High-risk operations, like a bank transaction or a change to a cloud database, stop at a hard checkpoint until a human approves them.
1.3 · Tools & docs
The tools behind each standard, with the official docs to read. Each one is tagged with the standard it supports and the phase of our build where we'll use it.
Loading tools…
Pillar 2 of 4
The agent projects and products people actually use, grouped by the kind of agent system each one is. Use it to see which patterns from Pillar 1 have reached production.
2.1 · Projects & products
What each one does and which category of agent system it represents. Open-source ones link to Pillar 3, where you can contribute to them. Status changes, like a product folded into another, are noted on the card.
Loading projects…
Pillar 3 of 4
The most-starred open-source agent projects, with what you need to start contributing: how active each one is, its good-first-issue list, and its contributing guide.
3.1 · Projects to contribute to
Ranked by GitHub stars, or sort by activity or by open beginner issues (each repo's own "good first issue" or "help wanted" label). Activity and issue counts are a snapshot; stars and last push load live from GitHub. Each one is tagged with the standard it shows best and the phase of our build where reading its code helps.
Pillar 4 of 4
What NJR and SA have built so far, measured against the standards in Pillar 1, and the experiments we're running along the way. The full tracker is the Repo progress page.
4.1 · Status
We're building every layer by hand: a neuron first, then an agent loop, then multiple agents, then the protocols and safeguards above. The full tracker has every phase, task, contribution and commit.
Open repo progress → See the execution plan →4.2 · R&D
The questions we're testing as we build, and the research behind them. Each experiment links to the papers it builds on and to the phase where we'll run it. The benchmarks are how the field measures agents, and how we'll measure ours.
Loading R&D…