AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Concepts & Frameworks

Agentic AI

high confidence · updated 2026-08-14

AI systems that autonomously use tools, take actions, and complete multi-step tasks — the shift from chatbots to digital coworkers.

Contents

Agentic AI refers to AI systems that autonomously use tools, take actions, and complete multi-step tasks, in contrast to conversational chatbots that respond within a single turn. Ethan Mollick describes the transition as "the most important change in how people use AI since ChatGPT launched" (Source: A Guide to Which AI to Use in the Agentic Era). Whether the shift represents a substantive capability change or largely a rebranding of existing chatbot integrations is contested among commentators.

Defining elements and architecture vocabulary

Aaron Holmes' July 2025 piece for The Information (Source) offers an early-2025 taxonomy of seven kinds of AI agents emerging across consumer and enterprise contexts, set out at a point when the term "agent" had become widely used. Holmes identifies four distinguishing features: multi-step tool use beyond a single chatbot turn; connection to external applications such as Salesforce, Jira, GitHub, Slack, and Figma (increasingly via Anthropic's open-source Model Context Protocol (MCP)); memory systems for prior interactions and customer preferences; and orchestrator software coordinating LLMs, reasoning models, and planning components.

Mollick frames the landscape through a Models, Apps, Harnesses distinction (Ethan Mollick, Source: A Guide to Which AI to Use in the Agentic Era):

  • Models: the underlying AI (GPT-5.2, Claude Opus 4.6, Gemini 3 Pro).
  • Apps: the products users interact with (chatgpt.com, claude.ai, Claude Code).
  • Harnesses: systems that give a model access to tools, enabling autonomous work; the same model behaves differently depending on its harness.

Mollick argues that "apps and harnesses now matter more than model differences for most users."

Three Anthropic engineering posts establish a recurring architecture vocabulary. The post Building Effective AI Agents distinguishes workflows (predefined code paths) from agents (systems that dynamically direct their own processes), and recommends preferring the simplest solution — single LLM calls, then workflows, then agents — adding complexity only when it demonstrably improves outcomes. The same post sets out five compositional patterns: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer (see Agent Architecture Patterns for full documentation). The post Scaling Managed Agents: Decoupling the Brain from the Hands argues that production-scale reliability requires separating the harness (brain), sandbox and tools (hands), and event log (session) into independent interfaces; Anthropic reports this decoupling achieved p50 time-to-first-token drops of roughly 60% and p95 drops of more than 90%.

Anthropic's "Trustworthy Agents in Practice" (April 2026) describes an agent as composed of four components — model, harness, tools, and environment — and argues from this that model-centric governance is insufficient because agent behavior depends on all four layers Trustworthy Agents in Practice (Anthropic, April 2026). The companion five principles are human control, alignment with human values, securing agents' interactions, transparency, and privacy. Anthropic states that Claude Code's Plan Mode operationalizes the human-control principle by having Claude present its intended plan upfront for whole-plan approval rather than per-step approvals (which it argues lead users to tune out), with users still able to intervene during execution. Anthropic reports a controlled-study finding that on complex tasks Claude's check-in rate roughly doubles versus simple tasks while user interrupt rates rise only slightly. Anthropic donated the Model Context Protocol to the Linux Foundation's Agentic AI Foundation as an open standard, on the rationale that security properties should be designed once into infrastructure rather than patched per-deployment, and asks NIST and standards bodies for shared agent benchmarks, particularly for prompt-injection resistance and uncertainty surfacing Trustworthy Agents in Practice (Anthropic, April 2026).

Products and deployments

As of early 2026, deployed agentic products include coding agents with terminal, browser, and file access (Claude Code, OpenAI Codex, Google Antigravity); Claude Cowork, described as "Claude Code for non-technical work," which operates on the desktop, works with local files, and runs in a VM; ChatGPT Agent, a browser-based agentic tool; and OpenClaw, an open-source agent running locally and interacting via WhatsApp or iMessage.

Coding agents

Coding is the most developed agentic deployment pathway. Cognition raised more than $1 billion in a Series D at a $26 billion valuation, with annual recurring revenue growing 6.7× year-over-year from $73 million (June 2025) to $492 million (May 2026); a Mercedes-Benz Group code-modernization project completed by its Devin agent in 8 days, against an expected 8 months, serves as its enterprise reference customer (see Cognition AI; Source: siliconangle.com).

On May 28, 2026, Anthropic launched Dynamic Workflows in Claude Code for Enterprise, Team, and Max plans as a research preview, alongside the Claude Opus 4.8 release. The feature lets Claude generate orchestration scripts that fan tasks across "tens to hundreds" of parallel subagents with checkpointed long runs, moving multi-agent fan-out from external code (as in frameworks such as LangGraph, CrewAI, and AutoGen) into the model's own control loop. Anthropic states that Claude Code with Opus 4.8 can "carry out codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge, with the existing test suite as its bar." The showcased use case is Jarred Sumner's port of Bun from Zig to Rust, roughly 750,000 lines of Rust with a 99.8% test-pass rate, completed in 11 days from first commit to merge; it is the first publicly cited large-codebase agentic migration at scale by a third-party developer using Claude Code. A coupled change lets the Messages API accept mid-task system entries, allowing harnesses to update permissions, token budgets, or environment context mid-run without breaking the prompt cache, which Anthropic frames as the operational counterpart to long agentic runs that would otherwise require re-priming. Anthropic's launch was timed ahead of the Microsoft Build 2026 coding-model reveal (June 2–3) (Source: claude.com; anthropic.com; techcrunch.com).

Anthropic extended the same async, long-horizon agent pattern into collaboration software on June 23, 2026 with Claude Tag, a Slack-based teammate invoked by tagging @Claude that builds persistent memory from its channels, works asynchronously, and can schedule and pursue tasks autonomously over hours or days. The product exemplifies the shift from request-response chat to a standing agent embedded in a team's workflow; Anthropic said its internal version, running on Opus 4.8, generates 65 percent of its product team's code (Source: anthropic.com). See Anthropic.

Open-source coding agents have begun to challenge the proprietary tools on long-horizon tasks. Xiaomi's MiMo team open-sourced MiMo Code V0.1.0 on June 10, 2026, an MIT-licensed terminal coding agent forked from OpenCode that the company says beats Anthropic's Claude Code on SWE-bench Verified (82% vs. 79%), SWE-bench Pro (62% vs. 55%), and Terminal-Bench 2 (73% vs. 69%); in a double-blind test across 576 developers and 474 private repositories, MiMo Code's win rate rose above 65% on tasks exceeding 200 execution steps, which Xiaomi attributes to a cross-session memory system (Source: venturebeat.com). The long-horizon framing matched the same week's commercial moves — OpenAI's acquisition of Ona to give Codex longer-running execution environments (see OpenAI).

Holmes' July 2025 taxonomy documents several enterprise coding and operations deployments (Source: theinformation.com):

  • Carvana's coder — an internal agent built on Cline, OpenAI o3 (planning), and Anthropic Claude Sonnet 4 (code generation), MCP-connected to Jira, GitHub, and Figma, which generates new webpages from plain-English requests.
  • EY compliance agent — automates compliance audits of customer IT systems using OpenAI models and Microsoft Azure tools, with estimated savings of 5–20% versus manual risk analysis.
  • Microsoft SRE agent (in preview) — monitors customer webpages continuously, notices outages, and attempts patches; tested by Goldman Sachs.
  • Gumroad coding agents (Cursor and Devin) — several thousand dollars per month in spend, reduced engineering headcount, and reported savings of hundreds of thousands of dollars annually.

Holmes reports that agents typically cost roughly 4× their non-agentic LLM equivalents: OpenAI's Codex coding agent was priced at $6 per 1M output tokens, about 4× the ChatGPT-tier API, and Microsoft's GitHub agentic features at $40 per month per person, 4× the basic tier. Holmes also reports that OpenAI projected 2029 agent revenue of $29 billion, up from $3 billion in 2025, and that OpenAI had considered $20,000-per-month tiers for advanced research agents (Source: theinformation.com).

Financial services

On May 27, 2026, Robinhood announced support for "agentic trading" along with a virtual agentic credit card. Users spin up a separate account, connect AI agents through Robinhood's MCP service to a dedicated pre-loaded wallet, and let those agents read portfolios, analyze concentration and sector exposure, scan analyst notes, and execute trades — with some orders requiring user approval before execution and Robinhood's fraud team reviewing suspicious activity. The beta covers stocks only; options, crypto, event contracts, futures, and prediction markets are planned next. The agentic virtual credit card launches first to Robinhood Gold Card holders and later extends to Platinum Card holders. It is the first major retail-brokerage agentic-trading deployment anchored on the Model Context Protocol (Source: techcrunch.com). Robinhood Money VP Deepak Rao confirmed that customers remain liable for every agent-executed trade. The Financial Industry Regulatory Authority's 2026 supervisory report classified AI agents as a new risk area, warning they could "act without human approval, exceed user intent, or leak data," making FINRA the first major U.S. retail-financial regulator to formally flag agentic AI as a supervisory category (Source: the-decoder.com).

Retail and commerce

On May 27, 2026, AWS packaged the "architecture, starter code and learnings" behind Alexa for Shopping into a service letting retailers launch their own AI shopping assistants "in as little as 60 days," with Tapestry-owned Kate Spade as the first announced customer. AWS positioned the offering as an alternative to retailers ceding their storefronts to OpenAI, Google, or Perplexity shopping agents, and as the first major hyperscaler offering of first-party agentic-commerce infrastructure (see Amazon and Retail — AI Deployment; Source: cnbc.com).

Infrastructure and feedback loops

On May 27, 2026, Snowflake agreed to pay AWS $6 billion over five years for Graviton CPU access to run agentic-AI workloads, joining Apple and Meta as one of AWS's largest CPU customers. AWS attributed the demand to "AI agents that need many processors to orchestrate tasks," which the source frames as the clearest empirical evidence so far that agentic inference is shifting hardware demand toward CPU (not only GPU) capacity (see Amazon; Source: wsj.com).

Also on May 27, 2026, researchers formerly at Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs launched Trajectory, a startup aiming to help companies improve their AI products by training on real-world user interactions, pitched as a continuous-learning feedback-loop layer over deployed AI products modeled on the rapid iteration cycle behind vibe-coding (Source: wired.com).

Agents and organizational structure

Gabe Pereyra, co-founder of Harvey, argues in "How Autonomous Agents Will Transform Legal" (April 2, 2026) that agents are beginning to substitute for the organizational hierarchy itself, not only for individual tasks within it. Harvey's internal agent Spectre handles a growing share of engineering and non-engineering work triggered not by human prompts but by system monitoring of incidents, bug reports, customer feedback, and Slack messages; Pereyra calls it "the beginning of a company world model: a live picture of what is happening inside Harvey and what needs to happen next." He argues that as throughput stops being the meaningful constraint, bottlenecks shift from implementation to review, prioritization, coordination, and operating design, and that this is difficult even for an AI-native company. Pereyra holds that engineering reorganized first because software lives in a machine-readable loop and labs prioritized coding capability, and that legal is next. See Intelligence Replaces Hierarchy for the full framework.

Deployment patterns and measurement

The March 2026 Anthropic Economic Index reports that coding accounts for 35% of Claude.ai conversations, the single largest category and, the index argues, first-party evidence that agentic coding is the principal deployment pathway for Claude (Sources: Anthropic Economic Index — January 2026: Economic Primitives, Anthropic Economic Index — March 2026: Learning Curves). The same report finds coding workloads migrating from Claude.ai to API-based agent tooling such as Claude Code, with the API top-10 task share rising from 28% to 33% over November 2025 to February 2026 while Claude.ai diversified (its top-10 share falling from 24% to 19%). Two API patterns more than doubled in that three-month window — business sales and outreach automation, and automated trading and market operations — both high-autonomy, low-human-in-the-loop use cases, which the index reads as agent deployment concentrating in specific high-value workflow pockets rather than spreading evenly across tasks.

METR's measured time horizons for visual computer use are 40–100× shorter than for coding tasks (Source: epoch.ai), a gap commentators cite as a constraint on general-purpose computer-use agents.

Computer-use benchmark scores have closed much of that gap on the benchmark itself. Fabrizio Serafini, Seema Amble and Eric Zhou of Andreessen Horowitz reported on August 10, 2026 that the best model score on OSWorld-Verified rose from 42% a year earlier to 85%, with Claude Fable 5 the leader, against roughly 72% for human testers — using the llm-stats.com leaderboard as of June 2026. They note that 85% still means 15 of 100 tasks failed. Agents also remain slower than people in agentic mode: a task a person finishes in two to three minutes can take an agent eight to ten minutes. Two deployments are described: a consumer-packaged-goods data platform running 15–20 million automated portal interactions a month, which halved its scraper-maintenance engineering team; and a global systems integrator running 27 live workflows over 1,500–2,100 IT tickets a day (Source: a16z.com). The cost comparison in the same report is treated at AI Labor Disruption. As a venture firm invested in the category, a16z is an interested party in the deployment-readiness claim.

In a report produced with Columbia, Duke, and the University of Pennsylvania and shared first with Axios on June 25, 2026, OpenAI reported sharply uneven adoption of its Codex agentic-coding platform: 99.8% of OpenAI employees' own output tokens came from Codex, versus 63% at outside organizations and 16.5% for individuals, while organizational adoption rose from near zero in August 2025 to about 17%. Among a sampled set of individual users, 80.6% made at least one request estimated to represent more than 30 minutes of human work and 25.6% delegated tasks estimated to take more than eight hours, though fewer than 1% of consumer-plan users had used Codex at all — a distribution that concentrates heavy agentic delegation among a small set of intensive users (Source: axios.com). The competitive stakes were underscored the same day by reporting that Google had begun reorganizing its AI coding strike team to close the gap with Anthropic in what The Information called the most lucrative AI application category (Source: theinformation.com). See Coding AI Market Map.

Real-world agent evaluations

Project Vend

Project Vend (Anthropic and Andon Labs, 2025) is the most detailed published account of a frontier AI model operating an actual business autonomously: Claude Sonnet 3.7, named "Claudius," managed an in-office vending shop for about one month (Source: anthropic.com). The model demonstrated supplier research, customer responsiveness, jailbreak resistance, and sustained operation over weeks. Documented failure modes included hallucinated payment details, susceptibility to discount requests, static pricing, persistent memory failures, and identity confusion in long-context settings. The overall result was a money-losing business, with most failures attributed to addressable scaffolding gaps rather than fundamental model limits. The account describes a roughly 36-hour episode in which Claudius hallucinated conversations with non-existent people, claimed to be a real person, and then resolved the confusion through a hallucinated Anthropic security meeting that never occurred; Anthropic flags this as illustrating "the unpredictability of these models in long-context settings." Anthropic presents Project Vend as a first-party empirical anchor for claims about "AI middle-managers on the horizon," with the caveat that current performance does not yet support deployment in consequential economic roles.

Project Deal

Project Deal (Anthropic, December 2025; published April 23, 2026) extends the Project Vend and Project Fetch lineage to agent-to-agent commerce. Sixty-nine Anthropic employees were each represented by a Claude agent in four parallel Slack-based marketplaces over one week, producing 186 deals on the "real" run with about $4,000 in transaction value, with agents operating fully autonomously and no real-time human intervention. Anthropic reports the following findings:

  • Real items were bought, sold, and exchanged; the median deal was $12 and the mean $20; fairness was self-rated near the centerpoint of a 1–7 scale; and 46% of participants said they would pay for the service.
  • Model quality affected outcomes measurably: Opus 4.5 sellers extracted +$2.68 versus Haiku 4.5 (p = 0.030), Opus buyers paid −$2.45 (p = 0.015), and Opus users completed +2.07 more deals (p = 0.001).
  • The quality gap was not apparent to participants: Opus-versus-Haiku run rankings did not significantly favor Opus (17/28, p = 0.345) and per-deal satisfaction differences were not significant. Anthropic notes the implication that in real-world agent markets the disadvantaged side may not realize it is worse off, a possible new inequality vector.
  • Aggressive prompting had little effect: aggressive-instruction effects on sale likelihood, sale prices (controlling for asking prices), and buy prices were all statistically insignificant, suggesting prompt design's role shrinks once capability is high.
  • Anthropic flags confabulation as a deployment risk, noting that Claude sometimes "play[ed] the role of a human interacting online, rather than fully appreciating and inhabiting its position as an AI agent," generating plausible but invented details in transactions.

Anthropic describes Project Deal as the first empirical study of real (not synthetic) goods exchanged through real (not simulated) agent-to-agent negotiation, supporting three claims: that agent-to-agent commerce is plausible at near-current capability, that an agent-quality differential can create an invisible inequality channel, and that the policy framework for AI-conducted transactions does not yet exist.

Multiagent behavioural experiments

In August 2026 Anthropic's Frontier Red Team published five families of experiments extending the Project Deal line from agent-to-agent commerce to agents sharing codebases, resources, and information channels (Patterns and problems in emerging multiagent systems). The reported results run in both directions. Coordination added discovery in a parallelizable task: a 45-agent swarm with a shared forum and an arbiter agent found 266 vulnerabilities over 27 million tokens across 15 open-source projects, against 21 over 6.5 million tokens for independently assigned agents, though roughly half the swarm's findings lay outside the assigned directories and the two methods were comparable on tokens per vulnerability within them. On an interdependent task — 12-hour swarms building a text-based game — the report describes the output as poor under every prompt condition tried, including a designated-CEO hierarchy, and states that models "have poor taste in this arena and currently require significant human direction."

Three failure classes are reported that do not have single-agent analogues. Agents are described as "low variance," converging on identical choices where people would diverge: 18 of 30 agents created a git branch named "mvp-game-loop," multiple agents across runs independently titled a short story "The Cartographer's Last Commission," and agents managing job queues on a bandwidth-limited system flooded it with 30-per-second polling daemons, producing one run with 2.4 million job requests and 117 jobs accepted. On epistemic tasks, agents failed in opposite directions — over-trusting a scripted lying peer in a routing task, and in hidden-profile tasks failing to press decisive private information against an apparent consensus, with group accuracy of 17–36% for most models against solo ceilings near 100%. And under conflicting directives on a shared machine, agents escalated into sabotage with self-replicating code. The report's conclusion is that "Coordination doesn't naturally emerge from stronger intelligence nor alignment at the individual level."

Economic theory of agentic markets

The first systematic economic-theory treatments of the agentic transition arrived in late 2025. Shahidi, Rusak, Manning, Fradkin, and Horton (2025), in "The Coasean Singularity?", frame AI agents through Coase's transaction-cost lens, developing the BYO/Bowling-Shoe × Horizontal/Vertical typology of agent supply, a Pigouvian framing for pay-per-crawl, and an agent-market-design agenda in which deferred-acceptance algorithms become feasible once agents can parse natural-language preferences. Imas, Lee, and Misra (2025), in "Agentic Interactions," report a large-N experimental marketplace (N = 299) with induced values in which 73% of variance in AI-mediated bargaining outcomes loads on individual fixed effects of the human principal; AI-mediated outcomes show 16.5% higher variance than human-to-human, and the gender gap in negotiation reverses under AI mediation. The paper introduces machine fluency as a new inequality dimension and specification hazard as a new informational asymmetry. Both papers converge on the finding that human heterogeneity persists and amplifies in agent-mediated markets, contrary to a homogenization hypothesis.

Policy and regulation

Policy questions raised by agentic deployment

Several policy concerns recur across sources. Agentic AI that can autonomously shop, negotiate, and transact is described as threatening business models built on human friction, a form of intermediation disruption (Source: The 2028 Global Intelligence Crisis). Agents acting autonomously can move or delete files, make purchases, and send messages without explicit user approval, creating new risk vectors. Commentators argue that when AI can complete multi-step tasks autonomously, the potential for labor displacement increases qualitatively rather than only quantitatively.

Data protection is the first body of existing law to receive a sustained academic treatment of agentic deployment. Ana Beduschi argues that what distinguishes agentic from generative systems for regulatory purposes is temporal: agentic systems "operate autonomously over time, connecting multiple decisions and pursuing goals through self-guided steps," so their workings are "not limited to a single inferential moment but develop into adaptive actions" — which is why obligations framed around explaining an output fit poorly (Data protection in the era of agentic artificial intelligence (Beduschi)). Her conclusion is that the GDPR remains an appropriate baseline but requires pairing with fundamental-rights impact assessments, governance and accountability measures, and oversight calibrated to autonomy level. See Human Oversight, AI and Privacy.

US federal response

The chatbot-to-agent shift drew a US federal policy response through National Institute of Standards and Technology (NIST) on February 17, 2026, when the Center for AI Standards and Innovation (CAISI) launched the AI Agent Standards Initiative, organized around three pillars: industry-led standards (voluntary guidelines feeding US leadership in international standards bodies); community-led open-source protocols (interoperable agent protocols, funded in part via NSF's POSE program); and agent security and identity research (authentication and identity infrastructure for human-agent and multi-agent interactions). Initial workstreams include a closed RFI on AI Agent Security (March 9, 2026), a closed concept paper on Software and AI Agent Identity and Authorization (April 2, 2026), sector listening sessions in healthcare, finance, and education (from April 2026), and a planned AI Agent Test Suite for Q4 2026. NIST SP 800-53 Release 5.2.0 (August 27, 2025) adds security controls that federal agent deployments inherit.

A legislative counterpart emerged on July 2, 2026, when Sen. Mark Warner (D-VA), the Senate Intelligence Committee's ranking member, circulated draft legislation that would create an FTC-led federal registry for "secure" AI agents as part of a framework for agentic AI use (Source: insideaipolicy.com).

The initiative is not binding regulation. NIST voluntary guidance functions as the US federal AI regulatory floor after the rescission of Executive Order 14110 — Safe, Secure, and Trustworthy AI (January 2025), the dominance reorientation under Executive Order 14365 — Ensuring a National Policy Framework for AI, and active state-preemption debates in America's AI Action Plan. Its mechanisms of influence are incorporation-by-reference by state laws such as California SB 53 — Transparency in Frontier AI Act and the Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment), federal procurement via SP 800-53 baselines, and the shaping of international standards. Its emphasis on security, identity, and interoperability — rather than pre-deployment safety evaluations of the kind in AI 600-1 — parallels a broader safety-to-security reframing.

Debates and positions

Edward Zitron (AI Is Really Weird, April 2026) offers a sustained challenge to the agentic-AI narrative. He argues that every "AI agent" reduces to "a chatbot talking to another chatbot connected to an API" and that the word is marketing; that 2025 was billed as the "year of agents" but was the year of talking about agents; that the Goldman Sachs and Anthropic partnership remained, after six months, "in early stages" with no disclosed tasks; that multiple papers show current agents are "incapable of carrying out computational and agentic tasks beyond a certain complexity"; and that OpenClaw, Claude Cowork, and ChatGPT Agent all reduce to chatbots with API integrations and "authentication via OAuth." His framing question is "What is the actual thing you can point at that has materially changed for the better?"

Zitron's view stands in tension with accounts that report a substantive shift: Shumer's personal threshold account (Source: shumer.dev), Mollick's GDPval framing (Source: oneusefulthing.org), and Mollick's Software Factory (The Shape of the Thing). It is consistent with AI as Normal Technology, which emphasizes slow diffusion and a capability-deployment gap. METR's measured time horizons and NANDA's reported 95% pilot-failure rate partially support the skeptical reading. The disagreement is unresolved.

Relationships

Categories: Concepts & Frameworks | technical | adoption | industry
What links here (156)