Twenty named AI specialists that plan, reason, and act across your discovery program — each with a defined role, its own tools, and hard limits on what it's allowed to do. And twenty is only the baseline: spin up new agents and skills whenever your science demands them.
Most “AI in the lab” is a single assistant behind a search box. Halffield ships a team. A coordinating agent triages the work and delegates to specialists — one reasons about target biology, another designs the statistics, another runs the data pipeline, another watches compliance. They collaborate, hand off to each other, and can take real actions on the platform — always behind an entitlements + human-confirm + audit gate.
Every agent is configurable: swap its model tier, set its memory scope, adjust its tool allowlist, and choose its autonomy — suggest, act-with-approval, or autonomous. Governance is a dial, not a leap of faith.
Agents can operate the control tower — abort a run, approve a proposal, make a go/no-go call, advance a phase — but cannot self-approve when strict mode is on.
Every agent action is written to a hash-chained audit trail.
~105 governed tools, each behind a typed contract and a per-agent permission scope.
Twenty is the baseline, not the ceiling — define new agents, roles, and skills (or clone and specialize existing ones) with no hard limit, each inheriting the same tools, governance, and audit.