Agentic AI has moved from “cool demo” to “someone has to run this in production.” The easiest way to stand out in 2026 is not saying “agents” on your CV. It’s having hands-on experience with the handful of Python tools companies keep converging on.
This week’s edition is a practical toolkit: what to learn, why it’s in demand, and how to build proof fast.
The demand signal
LangChain is a good proxy for where the market is heading. UK vacancy tracking shows permanent job ads citing LangChain jumped from 18 to 144 in the latest 6-month window (to 18 May 2026), and those LangChain job ads co-mention Python (78%), RAG (73%), plus LangGraph (35%) and LlamaIndex (37%).
So if you’re wondering “what skills should I pick up to be relevant,” the market is basically telling you.
The 5 tools I’d prioritise for Python engineers
1) LangGraph (agent orchestration with state)
If you’re building agents that need to survive retries, long workflows, or multi-step tool use, LangGraph is becoming a default choice. It just shipped LangGraph 1.2.0 this week.
It’s also huge in usage terms (downloads are very high).
Build proof in a weekend: a “support triage agent” that:
classifies a ticket
pulls context from a small knowledge base
proposes an action
escalates with a structured handoff
CV signal: “LangGraph stateful workflows + tool routing + checkpointing + tests.”
2) LlamaIndex (RAG and knowledge assistants)
LlamaIndex is the cleanest “data-to-agent” story in Python, and the main package released 0.14.22 this week.
It also has strong download volume.
Build proof in a weekend: a “policy copilot”:
ingests docs
answers questions with citations
logs retrieval quality (top-k hits, miss rate)
CV signal: “LlamaIndex ingestion + retrieval + citations + eval set.”
3) PydanticAI (typed agents, clean contracts)
PydanticAI is gaining traction because it makes agents feel like normal software: typed inputs/outputs, structured responses, clearer failure handling. It released v1.97.0 this week.
Build proof in a weekend: a “contract-first agent” where every tool call returns a typed model and every agent response is a schema, not free text.
CV signal: “Typed agent outputs, validation, structured tool contracts.”
4) DSPy (optimising prompts and pipelines like code)
DSPy is useful when you want to improve outputs systematically rather than tweaking prompts by vibes. Latest release is 3.2.1 this month.
It also has meaningful download volume.
Build proof in a weekend: take your RAG pipeline and use DSPy to tune:
query rewriting
answer format
retrieval settings
then measure improvements on a small test set.
CV signal: “DSPy pipeline optimisation with before/after eval metrics.”
5) MCP (Model Context Protocol) for tool integrations
MCP is increasingly the “standard interface” for plugging agents into tools and data sources. The Python SDK is active (recent release v1.27.1).
And there’s an official Python package on PyPI.
Build proof in a weekend: an MCP server that exposes:
a Postgres read-only query tool
a “search internal docs” tool
then connect it to your agent.
Important: MCP and agent tooling have had security scrutiny recently, so treat tool execution and inputs as untrusted and build guardrails.
CV signal: “MCP server/client, tool permissioning, sandboxing.”
Worth knowing if you’re targeting enterprise teams
Microsoft Agent Framework (Python)
Microsoft shipped Agent Framework 1.0 recently as a production-ready release for Python and .NET, with a clear enterprise positioning.
The Python package continues to ship updates (recent PyPI release).
If you’re applying to Microsoft-heavy orgs, this is a nice “I’m current” signal.
What hiring managers should screen for in “agentic” candidates
If you want fewer false positives, stop scoring “knows LangChain.” Score these instead:
State and failure handling: retries, idempotency, checkpointing
Tool contracts: typed outputs, validation, clean error envelopes
Evaluation: even a small test set, with measurable pass rates
Observability: traces, tool call logs, latency and cost tracking
Security posture: sandboxing, least privilege, safe tool execution
The tool choice matters less than whether the candidate can ship something reliable.
Quick Python watch from the last 7 days
LangGraph 1.2.0 (May 11)
LlamaIndex 0.14.22 (May 14)
PydanticAI 1.97.0 (May 15)
Ruff 0.15.13 (May 14)
uv 0.11.14 (May 12)
Job of the week
Senior or Principal Backend Engineer | B2B payments | Fully remote | £90k–£130k + discretionary bonus
A fully remote backend role at a B2B payments startup (30 people, engineering team of 10). The key requirement is payments and ledger experience (or adjacent). Python is highly desirable, but if you’re exceptional on the payments side they can be language-agnostic.
What they need
Payments domain depth: ledger, reconciliation, settlement, double-entry concepts
Strong backend fundamentals: reliability, data consistency, auditability
Good engineering hygiene: CI/CD, observability, production support mindset
If you want the full brief, message me and I’ll share details.
Outro
Agentic AI hiring is quickly becoming “normal software hiring with extra responsibilities”: state, evals, observability, and security. If you want to be in demand, pick one orchestration framework, one retrieval framework, and one evaluation habit, then ship a small project with measurable outcomes.
Will the market consolidate around one agent stack, or will it stay fragmented like web frameworks?
Hiring? Contact:
Josh Smith
LinkedIn: https://www.linkedin.com/in/python-recruitment/
Email: [email protected]
Phone: 01727 225 552