AI Tools Radar: September 2026

September 2026 did not bring one blockbuster model release. It brought infrastructure — marketplaces, agent harnesses, and enterprise grounding layers that make agents deployable outside demos. If you are interviewing this month, expect questions that reference MCP connectors, agent orchestration, and "what would you put in Autopilot versus Copilot Code." This radar reads like a front page: datelined items, short paragraphs, no hype adjectives without a name attached.

On the wire

27 September — Anthropic launches Claude Marketplace

Anthropic opened Claude Marketplace with more than 2,000 connectors, plugins, and pre-built agents. Partners named in the announcement include CrowdStrike, Cursor, Harvey, Legora, Lovable, and Snowflake. Developers build with MCP (Model Context Protocol) and Agent Skills — the same connector story Microsoft and OpenAI have been telling, now with a distribution storefront attached. For hiring, this matters because "we integrate with Claude" increasingly means "we list an MCP server on a marketplace," not "we pasted an API key into a script."

For practitioners, the shift is distribution and review burden, not a new protocol. You still implement OAuth scoping, idempotent tool handlers, and schema versioning — but you also ship a listing page, documentation that survives non-expert installers, and support playbooks when a customer enables your connector in a regulated tenant. Teams that treated MCP as an internal integration now inherit product-surface expectations: semver on tool contracts, changelogs end users can read, and telemetry that separates your bugs from model drift. Interview signal: explain how you would certify a connector before it appears beside CrowdStrike or Snowflake in a marketplace grid.

25 September — Microsoft announces the new Copilot

Corporate vice president Jared Spataro unveiled a Copilot reorganised around three capabilities. Home combines Chat and Cowork, with Office in Copilot embedding Word, Excel, and PowerPoint inside the experience. Code targets builders who want to assemble their own solutions on the same technology stack as GitHub Copilot. Autopilot is the headline: a persistent, proactive personal agent designed to keep working when you step away, entering private preview at the end of September.

Microsoft also previewed Today — a proactive command centre surfacing mail, calendar, Teams threads, meetings, and tasks — entering private preview in October. The subtext for candidates: Microsoft is betting that the winning interface is not a chat box but an always-on chief of staff. PM and solutions-architect interviews will probe whether you understand permission boundaries for such an agent.

Autopilot changes the unit of work from "one prompt, one reply" to "one goal, many steps across hours." Practitioners must design checkpointing: which actions require explicit approval, how partial progress is surfaced, and how you cancel or rewind when the user returns. Code mode matters for builders who outgrow templates — same stack as GitHub Copilot means your custom agent inherits enterprise auth and deployment paths your IT team already approved. If you have only shipped synchronous chat, your credible gap-closer is a story about async jobs with human gates — invoice reconciliation, ticket sweeps, migration dry-runs — even if Microsoft's brand names were not on the slide.

25 September — Microsoft Work IQ

Work IQ grounds Copilot and agents in Dynamics 365 and Power Platform business data. Preview begins 30 September 2026 with rollout through October. The package includes a Work IQ MCP Server, CLI, and coding-agent plugin so developers can pull CRM and ERP context into agent workflows without bespoke ETL for every tenant. If your resume mentions enterprise Copilot deployments, interviewers will ask how you prevent an agent from leaking one customer's record into another's thread — Work IQ is Microsoft's answer, and competitors are shipping parallels.

Grounding is where demo agents die in production. Without tenant-scoped context, an agent hallucinates plausible account numbers; with sloppy grounding, it retrieves the wrong tenant's rows. Work IQ pushes responsibility toward platform-enforced boundaries rather than every app team writing its own RAG pipeline over CRM exports. For engineers, the MCP Server and CLI mean you can prototype in a coding agent locally while security reviews a single connector pattern — faster iteration, but only if you understand Dynamics entity permissions well enough to explain why a tool call was denied.

25 September — n8n Agents

Workflow automation vendor n8n shipped n8n Agents in preview: describe an agent in natural language, attach a model plus tools and existing workflows, then trigger it from Slack, a schedule, or any flow via a new "Message an Agent" node. Available on n8n Cloud; self-hosted requires extra setup; Enterprise tier coming. This is the democratised face of the agent wave — less glamorous than Autopilot, but closer to how mid-market ops teams actually automate.

n8n Agents blurs the line between "workflow with an AI step" and "agent with a schedule." Practitioners gain speed — reuse existing nodes as tools, trigger from Slack without building a custom bot — but inherit workflow debt: versioned flows, secret rotation, and observability across branches a model may choose dynamically. Self-hosted gaps matter for regulated shops; preview on Cloud is where most teams will experiment first. In interviews, contrast n8n's event-driven graphs with Autopilot's user-centric persistence: both automate, but failure isolation and rollback stories differ.

10 September — OpenAI Agents API (public beta)

OpenAI released the Agents API in public beta — the same harness that powers Codex, plus an OpenAI-hosted sandbox, tool search, programmatic tool calling, multi-agent support, and no additional fees beyond tokens and tools. For engineers, the significance is standardisation: the loop of plan → tool call → observe → repeat is now a product surface, not a weekend script. Interview loops at OpenAI partners increasingly include "design an agent with three tools and a failure mode."

Tool search is the under-appreciated detail: agents with dozens of MCP tools cannot stuff every schema into context. Search selects a subset per turn, which changes how you author tools — atomic, well-described, idempotent — and how you test — ensure the right tool ranks first for ambiguous intents. The hosted sandbox moves security left for teams that previously ran arbitrary code on worker VMs with duct-taped isolation. Multi-agent support invites orchestration questions: supervisor versus peer agents, shared memory, and cost caps when sub-agents loop. Billing as tokens-plus-tools means finance will ask for per-workflow budgets; prepare a answer that mentions hard stops and downgrade paths.

Also on the radar

• TypeSafe AI / Jev — System One decision models in early access; see our deep dive on what Jev is and when it is not an LLM.
• Meituan — LongCat-2.5-Preview for long-horizon multimodal workflows.
• Teradata — Tera agentic coworker for analytics and data platforms.

Jev sits on the opposite end of September's spectrum from marketplace gloss: a typed decision layer rather than a conversational shell. Practitioners evaluating it should read the launch alongside OpenAI's harness — high-volume routing and guardrails versus language-heavy orchestration — and plan hybrid stacks accordingly. Meituan's LongCat preview targets long-horizon multimodal jobs; unless your interview is region-specific, one sentence on "multimodal agents need checkpointing and evals across modalities" suffices. Teradata's Tera names the analytics-platform pattern: agents that live next to warehouses and BI, with governance inherited from the data platform — familiar territory for data engineers suddenly asked agent questions in loops.

Launch summary table

Analysis: the common thread across September

Strip the press releases and three structural shifts repeat. First, agents moved from demo to platform: OpenAI productised the harness, Microsoft split Copilot into modes with distinct lifecycles, Anthropic attached a storefront. The engineering work is less "prompt the model" and more "operate a runtime" — sandboxes, connector auth, eval gates, cost telemetry. Second, MCP is becoming connective tissue rather than a Anthropic-only story. Marketplace listings, Work IQ servers, and coding-agent plugins all assume tools expose structured context the model did not memorize in training. Third, persistence and proactivity are the new frontier: Autopilot, Today, Teradata's Tera, and background n8n schedules share a lifecycle where the user is not staring at the chat window when work completes.

What did not change: models still hallucinate, tools still need least-privilege, and enterprises still buy outcomes not architecture diagrams. September's launches are wrappers that make failure cheaper to contain and success easier to distribute — not a repeal of basic software engineering.

For staff engineers, the month's lesson is portfolio thinking: one runtime (Agents API or Copilot stack), one distribution path (marketplace or internal registry), one grounding layer (Work IQ or your warehouse), and one typed triage layer where volume warrants it. Missing any leg shows up in interviews as hand-waving about "the AI" without naming who owns auth, evals, or spend caps.

What to watch in Q4 2026

• Autopilot and Today private previews expanding — expect post-mortems on permission prompts and user trust, not just feature lists.
• Work IQ rollout through October — watch for first public case studies on tenant isolation in agent threads.
• Marketplace certification requirements hardening as connector count grows — broken third-party tools become Anthropic's support burden.
• Agent eval tooling consolidating: golden tasks, regression suites, and LLM-as-judge patterns copied from CI into agent promotion workflows.
• Pricing pressure on background agents as finance teams see token burn from unattended loops.
• Competing marketplaces from other model vendors — distribution wars may matter more than marginal benchmark gains.

Three themes behind the headlines

First, MCP is becoming the USB-C of agent tooling — Anthropic, Microsoft, and the OpenAI ecosystem all gesture at connector standards rather than siloed plugins. Second, persistence beats chat: Autopilot, Today, and Teradata's Tera share a design language of background work with human checkpoints. Third, grounding in business data (Work IQ, Snowflake partnerships) separates demo agents from ones enterprises will pay for.

None of this eliminates LLMs. It wraps them in harnesses, sandboxes, and typed decision layers (see Jev) so agents fail less expensively when they are wrong.

How to talk about this in an interview

You do not need to name every September date. You need one crisp mapping from launch to failure mode. Example: "Claude Marketplace changes connector delivery — I would still enforce OAuth scopes server-side because the listing does not replace auth review." Example: "OpenAI Agents API standardises the loop — I would invest in tool search quality and sandbox egress rules before adding a fourth agent." That pattern beats reciting partner logos.

Sample ninety-second briefs

1. 1Marketplace: distribution for MCP connectors; engineering work is certification, semver, and support — not a new integration paradigm.
2. 2Autopilot: async agent with checkpoints; design permission boundaries and cancellation, not just prompts.
3. 3Agents API: harness + sandbox + tool search; treat tools like microservices with idempotency and observability.
4. 4n8n Agents: agents inside existing automation graphs; good for ops teams, different failure modes from user-facing copilots.
5. 5Work IQ: tenant-scoped CRM/ERP context; answer data-isolation questions with platform grounding, not ad-hoc RAG.

What this means for your next interview

Recruiters and hiring managers are updating question banks faster than blog posts. Expect behavioural variants of "tell me about a time you shipped an automation" to become "tell me about a time you shipped an agent" — with follow-ups on evals, tool permissions, and cost caps. System-design prompts that used to stop at "design a rate limiter" now add "and an agent that clears low-confidence tickets overnight."

1. 1Skim one announcement you have not read yet — Marketplace listing requirements, or OpenAI Agents API tool-search docs — and be ready to summarise it in ninety seconds.
2. 2Prepare one story where you constrained an LLM (schema, allow-list, human approval) even if you have not shipped Autopilot.
3. 3Know the difference between a workflow (n8n) and a persistent agent (Copilot Autopilot) — interviewers use them interchangeably to test precision.
4. 4If you mention Jev or System One models, pair them with LLM use cases; hybrid stacks read as mature.

September 2026 will look like a turning point in retrospect because distribution caught up with models. The candidates who sound current are not the ones who name every launch — they are the ones who explain which launch solves which failure mode in a production agent stack.

Keep a personal one-pager: five launches, one failure mode each, one story from your own work that rhymes. Update it when October previews land — Autopilot trust studies, Work IQ isolation write-ups — but do not replace fundamentals. Connectors still leak when auth is sloppy; background agents still need budgets; evals still beat demo magic. The radar is the front page; your interview answers are the editorial.