Bound the work
Set token and step allowances, refuse work that exceeds known limits, and share resource scopes across workers. Branches and delegated assignments carry their inherited constraints.
Software / Pollard
A runtime for agent work with explicit resource limits, approval gates, inspectable execution trees, and replay.
Agent infrastructure
Pollard records model and tool calls in a content-addressed execution tree. Keep your application’s model clients and agent loop while adding explicit controls at the call boundary.
Set token and step allowances, refuse work that exceeds known limits, and share resource scopes across workers. Branches and delegated assignments carry their inherited constraints.
Resolve tool calls against versioned action definitions. Policies can allow, refuse, or require approval; delegated permissions can narrow as tasks pass between agents.
Use stored semantic steps to audit a run and replay recorded results in tests. Branch and rollback operations preserve the recorded history while you explore alternatives.
Packages / November 13, 2026
Python, TypeScript, and Rust have independent release versions. The native TypeScript and Rust releases target Python 1.6.0 runtime behavior; Python 1.6.2 includes the optional Codex CLI adapter.
Install with pip install pollard. Includes agent teams, inherited scopes, checkpoints, durable approvals, and optional Codex CLI task integration. Python 3.10 or newer.
Install with npm install pollardai. The native Node.js runtime supports streaming, replay, persistent stores, shared resource limits, provider adapters, and audit tools.
Install with cargo add pollardai. The native Rust runtime includes eight storage backends, async streaming, shared resource limits, replay, and audit tools.
Build with Pollard
The Python quick start uses an ordinary local function, so you can inspect a governed run before connecting a model provider.
Install the Python package and follow the credential-free example. Create a runtime, set a token and step allowance, invoke a local model callable, and inspect the resulting node and usage report.
Build a team with recorded task identities and separate cursors. Delegate work with inherited resource limits and narrowed tool permissions, then inspect dependencies and results together.
Python 1.6.2 adds an optional adapter for Codex CLI tasks. Record the task inputs and returned answer, apply execution controls, and reuse the stored result through strict replay.
Research and materials
The paper connects agent control, audit records, and replay across frameworks.
Muntaser Syed and Marius Silaghi. Presented November 3, 2026 at IEEE ICTAI in Boca Raton. Preprint: arXiv:2610.16145.
Swarm Forest explores governed agent teams with shared limits and drift detection. ReleaseProof connects evaluation results, approval, execution controls, and drift holds in an agent-release workbench.
Talks and workshops
Conference sessions connect the runtime to governed agent workflows and open-source development.
A hands-on workshop at DevFest Tampa 2026: Build, Secure, Scale, at USF Tampa Campus, Engineering Building II. Exercises cover resource limits, action approval, evidence verification, and strict replay.
An accepted hands-on Python workshop at DevFest Miami 2026: AI + Art Edition, at Mana Tech Loft 10.
A practical session on packaging, testing, documenting, and releasing Python, JavaScript, and Rust libraries, illustrated with tokenmaster and Pollard.