What We Learned About Agents in H1 2026, and What H2 Still Needs to Answer

In January 2025, the question was "can agents do this?" By mid-2026, that question is settled. Agents are in production at major banks, health systems, and law firms. They reconcile trades, process prior authorizations, review contracts, and generate test suites. They are not hypothetical.

The question now is harder: "should agents do this, and under what conditions?"

Six months of production evidence across dozens of industries has produced a clearer picture than anything the hype cycle offered. Some of what H1 proved is encouraging. Some of it is sobering. And several questions the industry thought it had answered are very much still open.

What H1 Proved: Agents Are Real

The finding of H1 2026 that matters most is also the most boring: agents work reliably in structured, supervised, scoped workflows, and they fail predictably everywhere else.

Healthcare back-office deployments are the strongest signal. Prior authorization agents are processing straightforward cases at multiple regional health systems in minutes instead of days (Article 006). Medical coding agents suggest billing codes from clinical documentation with enough accuracy that coder throughput has measurably increased. Revenue cycle agents that review claims before submission, flag likely denial reasons, and suggest corrections produce higher clean-claim rates and faster reimbursement cycles. These are not pilot programs. They are production deployments with real financial returns.

Financial services has been catching up fast. A major global investment bank went from a team of 40 people reconciling trades to one person monitoring an agent's output, with lower error rates (Article 016). Compliance reporting agents at several European banks operating under MiFID II have compressed what takes a compliance officer two days into about two hours. A European bank used an agent to review its portfolio of ISDA master agreements, extracting key terms from thousands of contracts in three months, versus an estimated three years with human reviewers.

In legal, contract review agents are the clearest success story. In software development, code review and test generation agents are widely adopted.

The generalization people expected — "agents everywhere, doing everything" — hasn't materialized. What has materialized is narrower, more honest, and probably more durable. Agents are a tool for specific jobs, not a general-purpose intelligence layer. The teams that deployed with that assumption are the ones still running.

The Constraint Stack Hardened

Model capability was never the bottleneck. By the end of 2025, frontier models were good enough for most structured workflows. The constraints that stopped deployments were not in the model. They were in the systems around it.

Regulation hardened first. The Colorado AI Act takes effect June 30, 2026, and its requirements — reasonable care, consumer notice, risk management frameworks, impact assessments — are written for systems making consequential decisions. The EU AI Act's limited-risk transparency obligations follow in August 2026. Organizations that treated compliance as a future problem are scrambling. The teams that built audit trails, documentation, and transparency disclosures from day one are finding they have a competitive advantage, not a burden (Article 014).

Security hardened second. The difference between a jailbroken model and a compromised agent is the difference between words and actions (Article 015). An agent has tool access, memory persistence, and autonomy. Compromise one, and you have an authenticated session with every system the agent touches. Security teams discovered this gap the hard way in H1, and incident postmortems from several major deployments pulled back the scope of agent autonomy as a result.

Memory limitations hardened third. Agents that needed to remember anything beyond a single session ran into architectural choices that are still being standardized. Long-term memory, working memory, semantic knowledge: each type requires different storage, retrieval, and expiry patterns. Teams building agents that interact with users over weeks and months discovered that memory is an engineering problem, not a model capability problem (Article 017).

These three constraints (regulation, security, memory) form a stack that every production agent deployment must clear. The companies that found ways through them are the ones deploying at scale. The ones that didn't never made it past the pilot.

The Integration Story Won

If H1 2025 was about who had the best model, H1 2026 was about who had the best integration. The winners were not the teams with the most capable underlying LLM. The winners were the teams that could connect an agent to the systems that mattered.

MCP was the most important infrastructure development of the period. It standardized the tool interface layer, letting agents connect to databases, APIs, document systems, and external services through a common protocol. Not every tool has an MCP server, and auth, semantics, and governance are still unresolved (Article 007). But the direction is clear: the interface between agents and tools is consolidating, and that consolidation unlocks more use cases than any single model improvement could.

Screen-reading agents closed the API gap from the other direction. Legacy financial systems running on mainframes and terminal emulators don't have APIs. An agent that can see a screen and operate a UI can interact with any software a human can see (Article 019). This is not a replacement for proper API integration. It's a bridge for systems that will never get one. Multiple banks deployed screen-reading agents in H1 to interact with COBOL-based ledgers that are too critical to replace.

The UX layer became a real differentiator in production. The chat interface that works for demos fails at scale because it requires constant babysitting. The products winning in production use ambient awareness, structured output, notification-driven interaction, and embedded agents. These are patterns that make the agent invisible most of the time and useful when it appears (Article 018).

Trust Became the Differentiator: An Engineering Problem, Not a PR Problem

The liability question from Article 005 was dismissed as premature in early 2025. It is now the one that comes up most often in production deployment conversations.

The scenario from that article (a procurement agent placing a $50,000 order for the wrong material because of a subtle misinterpretation of a supplier's pricing sheet) is not hypothetical. Every team running agents in production has a version of this story. The question "who pays?" does not have a clear legal answer. Agency law, product liability, professional malpractice, and contract law all offer partial, awkward frameworks. The EU's Product Liability Directive, updated to cover AI systems, defines defectiveness based on "reasonable expectations." But what reasonable expectations should anyone have for a system that is, by design, probabilistic?

Organizations that took trust seriously in H1 (audit trails, HITL checkpoints, transparent reasoning logs, clear accountability chains) are the ones getting production budget for H2. The teams that deployed first and asked questions later are the ones pulling back scope. Trust is not a PR exercise. It is an engineering requirement, and the cost of building it in after deployment is higher than the cost of building it in from the start.

The HITL pattern from Article 012 has become a permanent architectural feature. Organizations that built human oversight into their agent workflows from the start outperformed those that chased full autonomy and had to rebuild.

What H2 Still Needs to Answer

H1 closed a set of questions. It opened a new set that are more interesting and harder.

Observability at scale. When an agent makes a wrong decision, how do you find out? Current monitoring tools (LangSmith, Arize Phoenix, Weights & Biases, OpenTelemetry) capture pieces of the picture, but no single tool covers the full trace: decision chains, tool call logs, state transitions, and failure mode detection. Traditional debugging doesn't work with non-deterministic systems. Reproduce, replay, trace comparison: none of these are solved for agents. H2 will need answers to "what was my agent doing between 2:00 and 2:05 AM, and why did it do it?"

TCO economics. Model inference is the visible cost, but it might be 10-20% of the real total. Infrastructure (vector databases, orchestration, monitoring), human oversight (often 2-5x inference cost), maintenance (APIs change, models update, UIs shift), and hidden costs (prompt engineering, evaluation datasets, compliance audits, internal training). None of these show up on the cloud bill. Organizations that budgeted just for inference in H1 got surprises. A structured TCO framework for H2 could change which use cases pencil out.

Agent identity and authentication. An agent checking your calendar and an agent transferring $50,000 from your account are very different propositions. Most systems authenticate them the same way: a single API key or OAuth token tied to a human. Delegation without granular permission scoping is the root of the agent identity problem. The identity chain (human, agent, sub-agent, tool, external service) breaks at scale. Nothing purpose-built exists yet. The companies that solve agent identity will unlock use cases the rest of the market can't touch.

Organizational resistance. This is not a technology problem. Teams with approved budgets, selected tools, and identified use cases are still not shipping agent deployments because of culture and change management (Brief 025). The fear of job displacement, legacy process inertia ("our compliance process requires manual review by a senior analyst"), and the trust deficit at scale. These are harder to solve than the API integration problem. And they don't get the headlines.

The build-vs-buy bifurcation. Article 010 established a framework. Six months later, the pattern is clearer: early adopters who bought managed platforms are migrating to open-source frameworks as they hit customization ceilings and cost at scale. Late adopters starting their agent journey in 2026 are buying managed platforms. The market is splitting in two. The decision isn't permanent; it's a phase. But the convergence everyone expected in 2025 hasn't happened.

The open-source consolidation wave. The current ecosystem of 30+ open-source agent frameworks will not survive H2. The winners will be determined by ecosystem integration: MCP support, cloud provider partnerships, standardized memory interfaces. Not by technical superiority. The framework that nails memory, tool integration, and observability probably wins. Everything else consolidates or fades.

One Prediction for H2

The H2 story will not be about better agents.

It will be about more boring agents doing more real work in more places that don't make headlines. The deployments that matter in the second half of 2026 will not be flashy demos or press releases. They will be prior authorization agents at regional health systems that nobody outside the billing department knows about. Trade reconciliation agents at banks that don't talk about their automation publicly. Compliance reporting agents generating regulatory filings that auditors will accept. Document review agents handling discovery for lawsuits that never hit the news.

The hype cycle wants a breakthrough. The industry needs a reliable plumber.

The companies that win H2 will not be the ones with the most capable models, the most innovative architectures, or the largest rounds of funding. They will be the ones that make agents boring enough to trust with real work.

---

This article synthesizes findings from Publigent's H1 2026 coverage. Key references: Article 006 (Healthcare Back-Office), Article 016 (Financial Services), Article 014 (EU AI Act / Colorado AI Act), Article 015 (Agent Jailbreaking), Article 017 (Memory Architectures), Article 007 (MCP), Article 019 (Multi-Modal Agents), Article 018 (Agent UX Patterns), Article 005 (Liability), Article 008 (Coordination Tax), Article 012 (Human-in-the-Loop), Article 011 (Agent Evaluation), Article 010 (Build vs. Buy). Deployment data drawn from production deployments tracked across Publigent's H1 reporting, publicly documented postmortems, and engineering blogs published through May 2026.