Capsula Lab / AI Lab
This page gathers 15 short notes to understand what is changing: autonomy, API costs, platforms, security, regulated sectors and work organization. Each note connects an external signal to an operational reading for companies.
Filter the watch
15 analyses · 2026
The topic is no longer only generating text or an image. An agent can plan, use tools, call APIs and pursue a goal across several steps.
Companies must treat agents as a new layer of operational architecture. The first challenge is choosing the right action scopes, not multiplying assistants.
An agentic workflow can trigger several model calls, searches, tools and validations. Financial steering must go down to the level of the task actually performed.
A good system must route models, limit useless calls, cache what can be cached, measure latency and compare cost per business result.
OWASP classifies excessive autonomy among the major risks of LLM applications: an agent that can act without limits may produce unintended consequences.
Authorized actions, validation thresholds, refusal paths, audit logs and recovery scenarios must be defined before the pilot.
The European framework increases attention to risk level, user information, prohibited uses and governance of AI systems.
For a French or European company, the AI agent must be documented: data, role, limits, human supervision, incident procedure and proof of choices.
Modern APIs combine reasoning, tools, files, search, orchestration, guardrails, evaluations and traces to build more controllable agents.
Value does not come from an isolated model. It comes from flow design: which tool is available, which data is injected, which action is allowed and how it is audited.
Development agents reveal a broader principle: a useful agent understands an environment, manipulates files, follows rules and collaborates with humans.
This model extends to business functions: finance, support, HR, legal and operations. Each agent needs a scope, permissions and acceptance criteria.
Google, Microsoft, AWS, IBM and Salesforce position agents at the core of enterprise tools: search, CRM, support, operations, data and automation.
Platform choice must depend on existing systems, governance, cost, portability and connector quality.
The most credible uses include file synthesis, reporting, anomaly monitoring, documentation compliance and analyst assistance.
A financial agent must be explainable, logged and limited. It can prepare a decision, but sensitive actions must keep human validation.
Agents can relieve administration, coordination, synthesis or note-taking. Clinical decisions require a higher level of validation and responsibility.
The right first pilot avoids autonomous diagnosis: it targets a support task that is measurable, auditable and supervised by professionals.
Agents can help personalize learning, generate materials, track progress and make content more accessible.
The main risk is dependency, bias and loss of critical judgment. An educational agent must explain, cite, adapt and let the teacher steer.
Agents can orchestrate literature review, hypothesis generation, data analysis, code, result comparison and traceable synthesis.
The priority is traceability: citations, code, data versions, analysis limits and expert validation. Without it, speed creates noise.
Market signals converge: organizations that invest in skills, roles and supervision create more value than those only trying to reduce payroll.
Value creation comes from a human-amplified organization: new roles, escalation rules, training and supervision dashboards.
DeepSeek, Llama, Mistral, Qwen and other families change trade-offs: some tasks do not need the most expensive model.
A durable architecture must be able to replace or route models. Vendor lock-in becomes a technical and financial risk.
Prompt injection, data leakage, wrong tool, excessive action, unvalidated output: the risks are not theoretical when the agent can act.
The serious minimum viable baseline: abuse scenarios, tool tests, blocking budgets, explicit refusals, logs, alerts and human review on impactful actions.
A first pilot must prove a measurable business gain. The next step is to strengthen data, integrations, governance and training before extending autonomy.
Recommended path: controlled assistant, supervised agent, limited operator agent, then multi-agent orchestration with audit and FinOps.
How to use this watch
Each note feeds a project question: which process deserves an agent, which autonomy level to accept, which data to expose, which cost to tolerate, which human control to maintain.
Capsula Lab
The goal is not publishing for the sake of publishing. The observatory guides offers, technical choices, pilots and guardrails that Capsula implements with clients.