AGENTIC AI · STATE OF THE ART · SEPTEMBER 2026
From generating answers to getting work done
The capabilities that define agentic AI today: orchestration, memory and reasoning, execution, and governance
Prashant Dhingra · DataKnobs
Generative AI answered prompts. Agentic AI pursues goals.
GENERATIVE AI · 2023–24
One prompt, one output
- Stateless: forgets once the session ends
- A human routes every step
- Output is a draft a person must act on
AGENTIC AI · 2026
Goal in, outcome out
- Runs a continuous perceive–plan–act loop
- Manages multi-step workflows for hours or days
- Acts on live systems within set permissions
Agentic AI capabilities · 2026
2
The control loop at the core of every agent
01
Perceive
Read the goal, environment, tool results and memory
02
Plan
Break the goal into steps and pick the next action
03
Act
Call a tool, run code, browse or delegate to a sub-agent
04
Reflect
Check the result, update memory, re-plan if needed
The loop repeats until the goal is met, a step or budget limit is reached, or a human is pulled in.
Agentic AI capabilities · 2026
3
Four capability layers define the modern agent stack
01
Multi-agent orchestration
Supervisors, shared context, long-horizon runs
02
Memory and reasoning
Persistent state, recursive inference, live grounding
03
Execution and tools
Terminal, browser and desktop control; typed tool contracts
04
Governance and security
Agent identity, circuit breakers, adaptive escalation
Agentic AI capabilities · 2026
4
01 · MULTI-AGENT ORCHESTRATION
A supervisor agent decomposes goals and delegates to specialists
- The supervisor splits a high-level goal into subtasks and assigns each one
- Specialists get narrow instructions, tools and permissions, so each works in a smaller, cleaner context
- Reviewer agents critique outputs before results are merged or shipped
Supervisor · orchestrator
Planner
Researcher
Executor
Reviewer
Sub-agents run in parallel, report back, get re-tasked
Agentic AI capabilities · 2026
5
01 · MULTI-AGENT ORCHESTRATION
Agents share context in real time and run for days
Real-time context sharing
Shared memory, message queues and standard protocols (MCP for tools, A2A for agent-to-agent) let agents work in parallel without a human routing messages.
Long-horizon autonomy
Asynchronous runs over hours or days, with checkpoints, resume and re-planning as conditions change.
Delegation goes mainstream
In September 2026 Anthropic folded Cowork into Claude, which now decides on its own whether a request needs a chat reply or a longer agentic task.
Agentic AI capabilities · 2026
6
02 · MEMORY AND REASONING
Agents keep state across sessions and learn from past runs
Working
The current task, plan and scratchpad
Episodic
Past runs, errors and the fixes that worked
Semantic
Facts, preferences, knowledge graphs
Procedural
Reusable skills and playbooks
67%
Alibaba says its new Agent Context layer, which adds real-time context and long-term memory, can cut token use by up to 67% in knowledge-heavy work (vendor claim, Sep 2026).
Agentic AI capabilities · 2026
7
02 · MEMORY AND REASONING
Recursive Language Models treat long context as something to explore
- The prompt is loaded as a variable in a Python REPL
- A root model writes code to search, slice and chunk it
- It recursively calls sub-model instances on the slices
- It combines their answers into the final result
MIT CSAIL reports inputs beyond 10M tokens at comparable or lower cost. On one 132K-token benchmark, an RLM using GPT-5-mini scored 64.9% vs 30.3% for GPT-5 called directly.
A useful distinction
RLMs are an inference strategy, not a new model class. Hidden, self-verifying reasoning comes from reasoning models with extended thinking. The two combine well: a reasoning model makes a stronger root.
Source: Zhang, Kraska & Khattab, arXiv:2512.24601
8
02 · MEMORY AND REASONING
Agents ground their reasoning in live data
Agentic retrieval
The agent decides when and what to look up (web, internal docs, knowledge graphs) and refines its queries as it goes.
Verify before acting
Facts are checked against systems of record before the agent writes, spends or sends anything.
The caveat
Grounding reduces hallucination; it does not remove it. Sources can be stale, wrong or carry prompt injections.
Agentic AI capabilities · 2026
9
03 · EXECUTION AND TOOLS
Agents operate software directly, like a human user
TERMINAL
Coding agents in the CLI
Tools like Claude Code and Codex edit repos, run tests and open pull requests from the command line.
BROWSER
AI browsers
ChatGPT Atlas, Perplexity Comet and browser agents navigate sites, fill forms and read unstructured pages.
DESKTOP AND DEVICE
OS-level agents
Meta's Muse agent now works with files and apps on macOS; Google's ADK for Kotlin 1.0 brings on-device agents to Android.
Agentic AI capabilities · 2026
10
03 · EXECUTION AND TOOLS
Typed tool contracts let agents change live systems
From suggestion to action
- Tools declared with typed schemas the agent can validate against
- MCP servers discovered and called at runtime
- Plans revised when a call fails or returns something unexpected
- Real side effects: refunds, database edits, deployments
NEW IN 2026
Agents that can pay
Mastercard Agent Pay integrations let agents complete purchases with user authorization, issuer controls and verifiable proof of intent. Payment processors are adopting the open Universal Commerce Protocol.
Agentic AI capabilities · 2026
11
04 · GOVERNANCE AND SECURITY
Agents get their own identities, permissions and registries
| Control | What it does | 2026 example |
|---|---|---|
| Agent identity | Non-human identities with scoped, short-lived tokens | HashiCorp Vault agentic IAM |
| Agent registry | One governed catalog of agents and tools | AWS Agent Registry |
| Gateway policy | Identity-aware rules over model and tool traffic | Kong AI Gateway 2.0 |
| RBAC per agent | Each agent gets only the tools its role needs | Scoped sub-agent toolsets |
Announcements: Aug 31 – Sep 1, 2026
12
04 · GOVERNANCE AND SECURITY
Bounded loops and calibrated escalation keep autonomy safe
Circuit breakers and loop bounds
- Maximum steps, time and spend per run
- Detection of repeated or oscillating actions
- Automatic halt when a policy threshold is crossed
- Sandboxes and dry runs for destructive actions
Adaptive escalation
- Agent estimates its own confidence
- Routine checks are verified by the agent itself
- Novel, irreversible or high-value actions go to a human
- People supervise the exceptions, not every output
Agentic AI capabilities · 2026
13
What this means for teams building agents
1
Design the whole harness (loop, memory, tools, limits), not just the prompt.
2
Start with a supervisor and a few narrow specialists; add agents only when it measurably helps.
3
Give every agent an identity and least-privilege tools from day one.
4
Budget autonomy with step and spend caps and clear escalation rules.
Agentic AI capabilities · 2026
14
Sources
- Zhang, Kraska & Khattab (MIT CSAIL), "Recursive Language Models," arXiv:2512.24601 (Dec 2025, rev. May 2026)
- Alibaba Apsara Conference: AgentCore, Agent Native Cloud, Agent Context (Sep 2026), via aiagentstore.ai
- Anthropic folds Cowork into Claude, launches Docs and Slides (Sep 17, 2026), via The Next Web / agentic.ai
- Google ADK for Kotlin 1.0; Meta Muse on macOS (Sep 2026), via aiagentsdirectory.com
- AWS Agent Registry (Aug 31, 2026); HashiCorp Vault agentic IAM and Kong AI Gateway 2.0 (Sep 1, 2026), via notes.brunopedro.com
- Mastercard Agent Pay / AgentCard; Worldline Universal Commerce Protocol support (Sep 2026), via aiagentstore.ai
- ChatGPT Atlas (Oct 2025) and Perplexity Comet (Jul 2025), via Wikipedia "AI browser"
Agentic AI capabilities · 2026
15