Agents

Autonomous agents that check their own work

Give an agent instructions and tools and let it decide what to do — but not blindly. Several models collaborate inside one agent: a primary model acts, others verify before anything ships. It remembers across runs, refuses to repeat past mistakes, and turns its best runs into reusable skills.

Two modes

Autonomous when you want reasoning, workflow when you want control

Most jobs want an agent that figures out the steps. Some want a fixed pipeline you can read line by line. almyty does both, in the same product.

Autonomous · the default

The model decides

Set instructions, attach tools, and the agent plans, calls tools, and composes an answer on its own. Several LLMs collaborate in one agent, output is verified before it ships, memory and constraints carry across runs, and strong runs become reusable skills.

Workflow · deterministic

You decide, exactly

When a task must run the same way every time, draw it as a graph: input, LLM call, tool call, condition, loop, parallel, merge, sub-agent. What runs is what you drew, validated before it ships.

Inside an autonomous agent

One agent, several models, real guardrails

Not a single prompt behind a chat box. An autonomous agent in almyty is a small system that reasons, checks itself, and gets better with use.

Multiple models in one agent

Run a fast model to act and stronger models to check, inside the same agent. Each role can use a different provider, your choice per agent.

Verify before it ships

A verification pass gates the final output, and can fire mid-run too, so a bad answer is caught and retried instead of returned.

Memory across runs

The agent keeps facts, preferences, and context between runs, so it stops relearning the same things every time.

Constraints from failures

When a run fails, the lesson becomes a constraint the agent must honor next time, so it does not repeat the same mistake.

Promote runs to skills

Turn a run that went well into a reusable skill the agent can reach for later, instead of re-deriving the steps from scratch.

Replay any run

Re-run a past execution to debug it or reproduce a result, with the full transcript of decisions and tool calls.

Call it like any model

However it is built, it answers on one endpoint

Autonomous or workflow, every agent is reachable through an OpenAI-compatible API. Point an existing SDK at it — change the base URL and the model name, keep the rest of your code.

call-agent.sh
# any OpenAI client works — just swap the base URL
curl https://api.almyty.com/agents/$ID/v1/chat/completions \
  -H "Authorization: Bearer $ALMYTY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "agent",
    "messages": [{"role":"user","content":"…"}]
  }'

Workflow mode

When you want a pipeline you can read

For deterministic, auditable work, build the agent as a graph. The same control-flow nodes you would otherwise hand-code, wired in a visual builder.

inputoutputllm_calltool_callconditiontransformloopparallelmergesub_agent
app.almyty.com
The almyty workflow builder with the node palette open — input, LLM call, tool call, condition, loop, parallel, merge, and sub-agent nodes on a pipeline canvas.

Branch on conditions

Route the run down different paths based on a tool result or an LLM decision, without leaving the graph.

Loop over data

Iterate a step across a list, with a guard so a run cannot spin forever.

Parallel and merge

Fan out independent calls, then merge their outputs back into one path.

Build your first agent

Start autonomous and let it reason, or draw a workflow when you need exact control. Connect an API first to give it tools to call.