The year agents started doing the work

2026 is the year finance stopped asking AI to search and started asking it to work. Charlie Flanagan, Chief AI Officer at Balyasny Asset Management, puts it plainly: the models changed, but so did the harness around them. Claude Code and similar execution layers let an AI take an outcome rather than a prompt, and keep working until the outcome is delivered.

The concrete example is merger-arbitrage analysis. When a deal is announced, an agent now builds the initial deal-analysis package. It estimates how likely the deal is to close and how long it will take, extracts the key economic and legal terms, identifies conditions and milestones, and flags the areas that need investor judgment. A year ago that work was fragmented across manual research and separate tools. It took three to five days. Now it takes under one day, and the agent itself runs for about 30 minutes before a human reviews anything material.

BAM is one node in a cluster of releases that all point the same direction. Anthropic shipped an open-source financial-services repo full of reference agents for investment banking, equity research, private equity, and fund operations. OpenAI launched Astra for Law and showed Cooley building an IPO workflow on ChatGPT Work. And a self-hosted stock monitor called PanWatch hit GitHub trending, wiring a nine-agent investment debate pipeline into a personal watchlist app.

Different scales, shared shift. Give the model an outcome instead of a prompt, let it work until done, and put a person at the review gate. The axis that separates these deployments is execution-layer ownership: a regulated institution running its own gatekeeping, a vendor workflow platform, or a developer's own server.

How Balyasny evaluates a frontier model before turning it on

BAM got here by building evaluation systems years ago, and that investment is doing the heavy lifting now. The firm tests new models on thousands of real financial tasks with verifiable outcomes across equities, macro, and commodities. General benchmarks and isolated demonstrations stay out of the decision.

BAM tests both the model on its own and the model inside its agentic environment, with the same tools, files, and requirements the firm's users have. Can it plan the work? Choose and use the right tools? Find and analyze evidence? Recover from errors? Check intermediate results? Produce a grounded deliverable? The failure modes BAM hunts for are specific: numerical errors, missed coverage, unsupported conclusions, retrieval problems.

On the relevant subset, Fable hit 89.4% versus 86.1% for the prior production model across thousands of tasks. A 3.3-point gap reads small on paper. In practice it decided which model a $38B firm routes production work through, because the score is measured on the firm's own tasks in the firm's own runtime.

The standout result was a set of economics problems that no model BAM had tested had ever completed. The team initially treated it as a broken evaluation, because the jump looked too large. They reran the test, independently checked the task and scoring logic, and reviewed the result with Anthropic before concluding the improvement was real. That rerun-then-verify loop is the part most teams skip, and it's the part that turns a wow moment into a routing decision.

Key numbers. BAM manages roughly $38 billion with about 2,000 investment professionals and support staff. Fable scored 89.4% versus 86.1% for the prior production model across thousands of real financial tasks. Merger-arbitrage packages went from three to five days down to under one day, with a 30-minute agent run and human review before material output. A tax-loss harvesting agent explored 90,000 database tables and produced a result that covered more ground than the traditional approach. The Chief Economist's central-bank analysis dropped from about two days to roughly 30 minutes. BAMAgent hosts thousands of autonomous agents running around the clock after six months of internal build.

Governance is a product and operating-model question

BAM's Chief AI Officer frames safety as a product question, and the firm's controls read like an infrastructure checklist: approved data boundaries, least-privilege access, tool-level permissions, logging and traceability, human review for material outputs, and escalation paths for edge cases. Adversarial and failure scenarios get tested before access broadens.

One principle from BAM's rollout stands out: a more capable model does not receive broader authority simply because it reasons better. Fable plans and analyzes more effectively than its predecessor. It still can only use the tools and data sources approved for that user and task, and it cannot grant itself more access. Investment judgment and accountability stay with people.

That posture is why BAM built BAMAgent, an internal platform for deploying agents into approved enterprise workflows. Six months in, it runs thousands of autonomous agents around the clock. The platform gives agents the tools they need and nothing more. Workflows include building and maintaining a company research package, preparing for an earnings or macro event, and turning new evidence into financial scenarios. The agent plans the work, uses approved internal systems, runs the analysis, checks intermediate outputs, and returns a research artifact a person can review.

Two agent results from inside BAMAgent show the pattern. One agent ran a tax-loss harvesting analysis that explored 90,000 database tables, found the relevant mutual fund holdings, and built its own weighting system. Another, configured by BAM's Chief Economist, cut a recurring central-bank analysis from roughly two days to about 30 minutes, with the economist keeping review and judgment. That time collapse is what a working agent stack looks like.

The open-source playbook in anthropics/financial-services

Anthropic's financial-services repo is the reference version of what BAM built in private, published for anyone to fork. It ships nine named agents, each as a self-contained Cowork plugin and as a Claude Managed Agent template you deploy through the /v1/agents API. Same system prompt, same skills, you choose where it runs.

Quick Take: The model gets the attention, but the eval set, the execution layer, and the human-review gate decide whether finance and legal AI ships.

The README is blunt about scope. These agents draft analyst work product: models, memos, research notes, reconciliations. They do not make investment recommendations, execute transactions, bind risk, post to a ledger, or approve onboarding. Every output is staged for human sign-off. That scope line is the governance model, shipped as documentation.

The nine named agents and the workflow each one owns:

Agent	Workflow it owns
Pitch Agent	Comps, precedents, LBO, then a branded pitch deck end to end
Market Researcher	Sector or theme to industry overview, market context, peer comps, ideas shortlist
Earnings Reviewer	Earnings call and filings to model update to note draft
Model Builder	DCF, LBO, 3-statement, comps, live in Excel
Valuation Reviewer	GP packages in, valuation template run, LP reporting staged
GL Reconciler	Finds breaks, traces root cause, routes for sign-off
Month-End Closer	Accruals, roll-forwards, variance commentary
Statement Auditor	Audits LP statements before distribution
KYC Screener	Parses onboarding docs, runs the rules engine, flags gaps

Underneath the agents sit vertical plugins with slash commands like /comps, /dcf, /earnings, and /ic-memo. The financial-analysis core plugin carries the shared modeling skills and all 11 data connectors, including MCP servers for FactSet, Morningstar, S&P Global, Moody's, LSEG, and PitchBook. If your firm runs Claude inside Excel, PowerPoint, Word, and Outlook through the Microsoft 365 add-in, the repo includes admin tooling to provision it against Vertex AI, Bedrock, or an internal LLM gateway instead of Anthropic's API. Everything is markdown and JSON, no build step, and a check.py script lints manifests and fails if a bundled skill drifts from its source. GitHub trending picked up both this repo and PanWatch in the same week, which says more about where practitioner energy sits than any launch post.

The legal vertical follows the same pattern

OpenAI's legal push uses the same playbook with different branding. Astra for Law brings frontier intelligence, custom firm workflows, connected legal data sources, and controls built for confidential client work. Cooley, a firm whose IPO practice spans decades of tech listings, built GO Public with ChatGPT Work to apply intelligence to the IPO process, helping lawyers surface issues earlier and keep judgment where it matters most.

The structure is the same shape as BAM's setup: a workflow platform, connected data sources, firm-specific process, and controls for regulated work. The IPO angle mirrors the merger-arbitrage angle, a high-stakes, document-heavy process where the first pass at issue-spotting is mechanical and the judgment layer is human. Public details are thin compared to the Anthropic repo, so treat Astra for Law as a signal rather than a spec. The signal is that legal workflows are now the same conversation as finance workflows.

The kitchen-table counterpoint: PanWatch

At the other end of the spectrum sits PanWatch, a self-hosted AI stock monitor that wires the TradingAgents framework, which sits north of 76,000 GitHub stars, into a personal watchlist app for A-shares, Hong Kong, and US stocks. No vendor platform, no enterprise controls, your data stays on your own box.

TradingAgents runs as a nine-agent pipeline. Four analyst types (technical, sentiment, news, fundamental) each read a position, then a bull-versus-bear debate happens, then a risk review, then a PM agent integrates a decision book. PanWatch exposes that as a single button on your holdings page. Click the brain icon, wait three to five minutes, and the full reasoning chain lands in Telegram, WeChat, or DingTalk.

The cost number changes the calculus: a run defaults to deepseek-chat and costs about $0.05. Run one analysis per position per day and you spend pocket change per month. That's the gap between enterprise AI and self-hosted AI in 2026. BAM spends on evaluation infrastructure and governance. PanWatch spends five cents and trusts the pipeline. The two approaches exist in different risk universes, and both are shipping.

When I set up PanWatch locally, the two defaults caught me off guard. The Docker image skips system-level Playwright dependencies, but Chromium still downloads into the mounted volume on first startup, which takes minutes and needs network access; set PLAYWRIGHT_SKIP_BROWSER_INSTALL=1 if you never need screenshots. And OpenTelemetry tracing is a no-op unless you set an OTLP endpoint, so the observability layer costs nothing until you opt in. When you do, agent runs map to root spans and individual LLM calls map to gen_ai child spans following the OpenTelemetry GenAI semantic conventions, which means you can push traces to Jaeger or Langfuse without vendor lock-in. Even a five-cent hobby agent ships with structured trace IDs and agent run tables these days. The bar for observability has moved.

What trips people up

These are the failure modes that show up repeatedly across this cluster:

• Widening permissions when you upgrade the model. BAM's rule is direct: a more capable model does not receive broader authority because it reasons better. When you move from your current model to a frontier one, re-run the eval suite and re-check least-privilege access rather than granting the new model more tools.
• Evaluating on public benchmarks. A model can ace MMLU and still fail on numerical errors, missed coverage, unsupported conclusions, and retrieval problems inside your stack. Build a task set from your own workflows with verifiable outcomes, and test the model with the same tools, files, and requirements your users have.
• Shipping agents without a review gate. The financial-services repo stages every output for human sign-off and explicitly forbids agents from executing transactions or binding risk. The moment an agent can act on its own output without a reviewer, you have changed the risk profile of the system. Keep a person on material outputs and log everything.
• Using the reference repo unchanged. The agents are starting points, and the README says so. Swap the MCP connectors to point at your data providers, drop firm terminology and formatting standards into the skills, teach Claude your PowerPoint layouts. A canned comps skill produces canned output; tuning is the work.
• Skipping the failure-scenario tests. You have to know how you'll detect a failure before you broaden access. That means logging agent runs, tracing decisions back to evidence, and having an escalation path for edge cases, the same way BAM tests adversarial scenarios before widening the rollout.

One thing to remember

If one idea from this cluster sticks, make it the eval-first posture. Balyasny invested in serious evaluation systems years before Fable existed, and that is why it could adopt a frontier model within days, route it against cost and efficiency, and recognize a never-before-solved economics problem as a real step change rather than a fluke. The model is a commodity input. The task set, the execution layer, and the review gate are the durable assets. Firms that build those will adopt each new model faster, and with less risk, than firms that start from the model.

The bottom line

If you run a regulated investment or legal operation, build your own eval set against real workflows before adopting a frontier model. BAM's 3.3-point gap between Fable and its predecessor only mattered because it was measured on thousands of the firm's own tasks with verifiable outcomes; on general benchmarks, that routing decision would have been guesswork.

If you are a solo developer or small team that needs AI research without enterprise overhead, self-host PanWatch or fork the financial-services repo. A nine-agent debate pipeline costs about five cents per run on deepseek-chat, your holdings never leave your disk, and the OpenTelemetry tracing stays off until you flip it on.

One thing to watch: the build-versus-buy split is still open. BAM built BAMAgent internally. Cooley built GO Public on a vendor platform. Anthropic and OpenAI both ship reference architectures that let you do either. Vendor platforms will keep absorbing more of the execution layer, and the open-source layer will keep undercutting them on price. Within six months, the model you run will matter less than the eval set, the data connectors, and the human-review controls sitting around it.