Model‑neutral agent platform for the enterprise

Developers build, validate and publish AI agents in the built‑in development environment. They also describe when to use each agent in plain English.

Users just chat with the platform. Semantic search routes each request to the agent whose description fits. The agent runs in a controlled runtime. Multiple agents build on each other in a conversation.

The platform handles all the plumbing and records every agent run for observability.

Developers
Users
Tactet agent platformpip install tactet
Agents
Build, validate, publish agents
Is part 7731 in stock in Berlin?Inventory: Berlin has 42 units of part 7731, below its reorder point of 60.
Describe agents in EnglishInventory: Reports stock levels by part and site.Purchasing: Raises purchase orders with approved suppliers.
Reorder the part.Purchasing: Done. PO‑5519 orders 18 units of part 7731 for Berlin.
Enterprise systems
LLMs

Demo: Chat with agents

An agent framework built for AI‑assisted development

An agent is written in JSON with Python expressions. Systems and prompts are defined outside the agent’s code: APIs, databases, MCP servers and LLMs in platform-wide profiles, and prompts in each agent’s meta.

The UI includes a complete development environment for agents. Tell Draft with AI what the agent should do, refine the generated agent in the schema‑aware code editor, then trace and validate it before publishing.

Agent structure
{ "functions": { "main": { "commands": { "api": { … }, "var": { … }, "llm": { … } } }, "order_status": { "commands": { "db": { … }, "return": { … } } } } } Agent: a set of functions Function. Every run starts in main Commands, run in this order Another function Its commands, in order
Commands
Variablesvar
LLM interactionllmmemory
External systemsapidb
Human-in-the-loopask
Agent outputoutput
Agent terminationendfail
Observabilityprogresstracelog

Demo: Draft an agent with AI

Conversations let agents build on each other

Users hold conversations from the UI, or from Slack and Teams with a bot. The LLM message thread is stored in a model-neutral format, so the next agent can continue the same thread with a different model. Conversation variables give agents short-term memory, and the memory command gives them long-term memory.

Conversation
Book my trip for Tuesday's all‑day meeting in New York
Travel desk UA1191 departs 06:00 Tuesday, back at 19:30, and a night at the Marriott near the office. $1,240. Book it? BookHold Submit
Travel deskBooked UA1191 and the hotel night. The confirmation is in your inbox.
File the expenses for that trip
ReimbursementsFiled the New York trip, $1,240, against your cost center. Sent to your manager for approval.

Demo: Trace an agent run

Agent harness as the orchestrator

Conversations are one way to use Tactet agents. An agent harness, such as Claude Code, is another. Tactet‑skill connects the harness to every published agent. The harness splits a request into parts, runs Tactet agents alongside its local skills, and combines the results. Each agent runs on the agent server, inside your organization.

Agent harness without Tactet

Harness context Many skillsMCP tool definitionsGenerated workflowIntermediate resultsResult
  • Accuracy and cost
  • Tool selection accuracy degrades sharply as the number of tools grows
  • Context rot from tool schemas, generated workflows and intermediate results
  • Tokens are spent on tool selection and generating the workflow
  • Every step runs on the same model
  • Security and maintenance
  • The model decides which enterprise systems are called at run time
  • Credentials for enterprise systems are included on each user’s machine
  • Session telemetry needs to be recorded separately
  • Every skill needs a retest when the harness model changes

Agent harness with Tactet

Harness context Tactet skillResult
  • Accuracy and cost
  • Search accuracy does not degrade with the number of agents
  • The harness context holds only the end result
  • LLMs are called only where the task needs them
  • Each agent uses the best or the cheapest model for its task
  • Security and maintenance
  • The enterprise systems called by an agent are defined at the time it is published
  • Credentials are managed centrally in a secret vault
  • Every agent run is recorded, including its LLM cost
  • A single Tactet skill needs a retest when the harness model changes

How the platform fits in your organization

The platform has three parts: a stateless agent server and a UI that run in your organization, and a cloud service that provides state, search and reporting. pip installs the agent server and the UI together. Each can also be installed separately from the source. The agent server can run standalone or in a cluster. The cloud service can be shared or dedicated.

Built for enterprise operations

Security & privacy

Secrets live in your vault. Conversations and memories are encrypted, by default or, optionally, end to end. Data privacy lists how each kind of data is handled.

Source available

Everything that runs in your organization can be built from the source: the agent server and the UI. Set the logging configuration to see every call to the cloud service.

Multiple environments

An agent refers to enterprise systems by profile name. Dev and prod profiles point at their own systems, so the same agent runs unchanged.

Horizontally scalable

Agent servers are stateless and can be run in a cluster. Configuration and profile changes are live on the next agent run on every server, without a restart.

Integration

Integrate at three levels: from Python expressions inside an agent, through the REST API, and with hooks that run your code at the platform’s decision points.