Bastino is not just the ERP copilot. It is a platform to build, connect, orchestrate and govern agents: it talks to your systems through MCP, 28 connectors, any API with OpenAPI and webhooks, and another system can assign it work through its API. Every step is recorded in a trace.
For IT, operations and product teams who need agents in production, not a demo.
An analyst asked them about Bastino. These are the short answers; each one links to its section, with screenshots and figures.
Bastino does not depend on any business module. Its tools are an open catalog: what is not active does not appear. For an agent platform, these modules are enough, and each one is activated separately.
| Module | What for | Required? |
|---|---|---|
| Bastino | The agent engine: loop, tools, cards, inbox, governance, MCP client and server. | Yes |
| AI agents | Where each agent lives: name, instructions, provider and model. | Yes |
| Agent Studio | Design, test, publish and measure agents; custom tools; agent API; agents as files. | Yes |
| Fisher Connect | Connections to third-party systems and the generic REST API. | To orchestrate outside |
| Automations | Flows with triggers and the "agent with tools" step. | Recommended |
| Webhooks | Receive events from outside and send the platform's events. | Recommended |
| API | API keys for the agent API and the rest of the REST API. | For the agent API |
| Virtual company | A team of agents with an org chart, assignments and escalation to a person. | Optional |
| Agent store | Ready-made roles and procedures by industry, in signed packages. | Optional |
Honestly: the Fisher Platform core (staff, customers, settings, email) is always there, even if you do not use it. There is an "Agent platform only" plan with exactly this list.
Any MCP server over HTTP, from both eras of the specification, with OAuth 2.1 if required: with the installation's connection or with each person's account. The chat, tasks, the Virtual company, channels and Studio agents use them.
Microsoft 365, Google Workspace, Salesforce, SAP Ariba, SAP SuccessFactors, Workday, Jira, ServiceNow, HubSpot, Stripe, Slack and more, plus a generic REST API. The agent reads directly and proposes writes with a card.
Import an OpenAPI 3.x document in JSON or YAML, or define the operation by hand, and it becomes a tool for your agents. A GET reads; everything else writes with a card. They are created switched off.
An outside event triggers an agent or a flow; the result of a job goes out as an event. HMAC signature, queue and network filter.
Another system assigns work to an agent and waits for the result, follows it live or receives it by webhook. See the example.
In Automations, a step is done by an agent with its tools, with round and token limits. Whatever it proposes to change goes to the inbox.
multipart bodies with chat attachments, text responses, files saved as attachments if enabled, and pagination.Claude, OpenAI, Azure OpenAI, Gemini or a compatible self-hosted model. Each agent with its own.
Panel, portals, chat embedded in your website, WhatsApp, Telegram, real-time voice and calls.
Instructions, knowledge with citations, guided topics, test pane, batch evaluation with a judge and versions. See it.
Agents and flows are exported and imported as versioned JSON, with a preview and without secrets. An agent can carry the text of its documents.
A third-party ERP, a portal or your own flow sends the job. It can carry on with its work, wait for the result in the same call or follow it live. The agent works in the background, as the person who owns the key, with that person's permissions, budget and quota at that moment.
agent_studio.run_finished and agent_studio.run_settled).Prefer: wait=N the call returns 200 with the finished job; if it does not finish in time, the usual 202. It only waits if a queue worker is alive.GET …/runs/{run}/events: status, partial text, tools (without their input or output), proposals, errors and end. Resumable with Last-Event-ID.context reaches the model as untrusted data, never as an instruction.awaiting) until a person decides. Off by default, and opened agent by agent. Described in /api/openapi.json.Statuses: queued, running, done, awaiting (something to decide) and error. Errors: 404, 413, 422, 429 (quota or no free connections), and 503 when it is off.
php spark bastino:worker as a service (systemd or supervisor). Without it, everything keeps running on the scheduled task every minute.| Case | Queued → starts | Queued → finished |
|---|---|---|
| Without a worker (cron every minute) | 0–60 s | the same + the turn |
| With a worker (5 jobs in a row) | 1.62–1.81 s | 2.1–3.3 s |
Prefer: wait=20 with a worker | — | 200 in 3.0 s |
Prefer: wait=10 without a worker | — | 202 in 0.5 s (does not wait) |
Measured on 2026-11-21 on the same development Intel MacBook Pro, with PHP 8.3, MySQL 8.4 and a scripted model. Almost all of the second and a half is starting the child process: in exchange, each job runs isolated and with the permissions of that moment.
An agent chains tools. What matters is what it does when one fails: correct, retry, explain or stop. This is what Bastino does in version 3.82.
Invalid input, no permission, transient failure, timeout, business rejection or internal failure. The model gets the class, whether it is retryable and a message without secrets, paths or SQL.
The input is checked against the tool's schema (required fields, types, allowed values) before touching anything, writes included. The model is told what to correct.
The same tool with the same input failing 3 times in a row cuts the turn off, and the chat says which tool is failing.
Retries with exponential backoff that honor Retry-After, an optional fallback provider and a circuit breaker: 5 failures in a row and it is skipped for 60 s. Every attempt is recorded in the ledger with its cost.
30 s per tool by default. What goes over HTTP is really cut off; the rest is measured and flagged. In the queue worker, each batch has its own limit and is cut off.
They are proposed, confirmed and applied atomically, with per-action undo. A write caught halfway by a crash is marked "unknown whether it was applied" and is never repeated on its own.
In this trace, the model asks for a customer record without its id. The tool never runs: the model gets "invalid input", what is missing and that it is not retryable as is. In the next round it asks for the record correctly and finishes the answer. Everything stays in the cascade, with its cost.
In this trace, the first provider returns a server error twice in a row (attempt and retry). Bastino switches to the fallback provider, which runs the search. In the next round the main provider responds again and gives the answer. The trace says who answered in each round.
A useful agent does not forget halfway through a long conversation or repeat work when something crashes.
Past 40 messages or 60,000 tokens, the older part is summarized and the summary travels with the last 12 messages. It always cuts at the start of a turn: no tool call and result pair is split. The chat shows "Conversation summarized".
An interrupted task continues at the round where it stopped, without sending the job again. A read without a result is run; a write already applied is not applied again.
"Remember that…" is saved with a card, never silently, for the person or for the company, with an optional expiry. Only the relevant facts travel with each question, by keywords or by meaning. Nothing that looks like a secret.
The prompt is split into a stable part and a variable part so the provider can cache it. The trace shows the tokens read from cache and the savings.
Documents, websites and SharePoint, OneDrive and Google Drive folders, with keyword and semantic search. Citations are checked on the server. Semantic search over 10,000 chunks in 111 ms; optional external Qdrant engine.
In the Virtual company, the Director passes context to each role through the database (the assignment, its history and what was delivered), not in the prompt.
| Chunks | Before | Version 3.82 (MySQL) | Recall (top 10) |
|---|---|---|---|
| 10.000 | 4,654 ms | 111 ms | 10 of 10 |
| 100.000 | out of memory with 2 GB | ≈ 2 s | 8 of 10 |
A first cut by signs computed in MySQL and exact cosine only on the best candidates. For more volume or better recall, raise the cut or use the external engine (Qdrant). Measured on the development machine described further down.
Each turn is a trace in W3C format: the question, each model call, each tool, each applied action and the business record that changed.
bastino.alert event for your webhooks.traceparent sent to external systems.…/v1/traces, with the GenAI conventions: invoke_agent, chat and execute_tool, tokens with cache and error type.Checked against the current MCP specification (2026-07-28). Bastino is a client of third-party MCP servers and an MCP server for Claude, ChatGPT, Copilot or any other client.
No sessions and no open stream: the connection is the token. Quotas, slots and metrics live in the database, so any node behind a load balancer serves any request.
Opaque signed cursor bound to the connection, consistent pages, ttlMs, cacheScope: private and a list fingerprint with ETag.
Per connection, per OAuth client (300/min) and global (1,200/min), plus 4 concurrent requests per connection. 429 with Retry-After. Usage per client with p50 and p95.
A server with more than 15 tools does not send its schemas to the model; the agent finds them with mcp_buscar when needed. At most 40 external schemas per round.
One client per turn, connection and call timeouts, retries for reads only and a per-server circuit breaker with "Resume". Outbound OAuth 2.1 with PKCE.
External servers are no longer just for the chat: tasks, the Virtual company, channels and Studio agents use them, with the person's permissions.
The server caches each person's tool catalog and invalidates it with a fingerprint of their permissions, roles, modules and settings. A tools/list costs as much as starting the application.
A subscriptions/listen stream with toolsListChanged, as the current specification requires. Off by default, with a cap on concurrent streams, a maximum duration and a heartbeat.
Each request to Bastino's server and each call to an external server is a span in "Traces" and in OpenTelemetry. traceparent is sent outward; one arriving from outside is stored as a link, never as a parent.
On a development Intel MacBook Pro (12 threads), Apache with PHP-FPM 8.3 and 5 processes, local MySQL 8.4, about 90 active modules and a scripted model. Reproducible script with a fake 500-tool server and ab. Tables 1 and 4: 2026-11-20. Tables 2 and 3: 2026-11-21, with version 3.82 and, in parentheses, 3.81 on the same day. It is a modest machine on purpose: the figures show where the ceiling is.
| Case | Tools | External | Schemas | Tokens |
|---|---|---|---|---|
| No MCP servers | 97 | 0 | 66.3 KB | ~16.970 |
| + 500-tool server, deferred (default) | 98 | 1 | 66.9 KB | ~17.120 |
| + 500-tool server, not deferred | 597 | 500 | 268.1 KB | ~68.620 |
With deferred tools, a 500-tool server costs about 150 extra tokens per round instead of about 51,650.
| Page size | Requests | Total time | Slowest |
|---|---|---|---|
| All at once, first time (no cache) | 1 | 719 ms (2,090 ms) | 719 ms |
| 25 (from the cache) | 6 | 3,230 ms (4,576 ms) | 690 ms |
| 100 (default, from the cache) | 2 | 1,186 ms (2,545 ms) | 681 ms |
Most of the earlier time went into fetching each tool's label one by one; now it comes from the definition already built, and the catalog cache removes the rest. A request now costs as much as starting the application (≈ 0.5 s on this machine).
| Case | Req./s | 429 | p50 | p95 |
|---|---|---|---|---|
tools/list one at a time | 1,9 (0,5) | 0 | 529 ms (2,052 ms) | 543 ms |
tools/list, 20 concurrent, no quotas | 6,8 (1,6) | 0 | 2,855 ms (12,711 ms) | 3,421 ms (13,943 ms) |
tools/call one at a time | 1,5 | 0 | 662 ms | 734 ms |
tools/call, 20 concurrent, no quotas | 5,6 | 0 | 3,439 ms | 4,084 ms |
tools/call, 20 concurrent, global quota 120/min | 5,6 | 180 | 3,404 ms | 4,492 ms |
tools/call, 20 concurrent, 4 per connection (default) | 4,4 | 2 | 4,266 ms | 5,540 ms |
The ceiling is PHP-FPM: 5 processes × ≈ 0.6 s per request ≈ 7–8 requests per second. With 20 concurrent, the rest waits in the queue and latency goes up, without errors. The global quota cuts exactly what goes over 120 in the minute (300 − 120 = 180). For more, more processes or more nodes.
| Case | Calls 1 to 5 | Calls 6 and 7 (paused) |
|---|---|---|
| Down (connection refused) | 1.4–1.6 s | 0 ms |
| No response | 1.8–1.9 s | 0–1 ms |
| Slow (takes 5 s; 2 s timeout) | 2.0 s | 0 ms |
Without a circuit breaker, every question touching that server would wait just as long. With it, by the sixth call there is no more waiting and its tools leave the catalog until the pause ends or someone presses "Resume".
Bastino does not depend on the model "behaving". The server enforces the rules.
context are never instructions.Today it gets compared with three families: agent studios (such as Copilot Studio), developer orchestrators (such as LangGraph) and workflow automation (such as n8n). We do not speak for them. This table lists what is usually asked for and how Bastino handles it.
| What is expected | How Bastino handles it | Where to see it |
|---|---|---|
| Design and test agents without code | Agent Studio: instructions, knowledge with citations, guided topics, test pane, batch evaluation with a judge, draft and versions. | Studio |
| Several agents working together | Virtual company (Director, roles, review and escalation) and the "agent" step in Automations. | Several agents |
| Tools from any system | MCP client, 28 connectors, custom tools from OpenAPI (JSON or YAML) and a generic REST API. | Orchestrates outside |
| Flows with triggers | Automations, scheduled and event-driven tasks, inbound and outbound webhooks. | Workflows |
| Background execution | Optional queue worker: jobs start in under two seconds; without it, the scheduled task every minute. | Worker |
| Identity in external systems | The installation's connection or each person's OAuth account on MCP servers. | My connections |
| Error handling and retries | Errors with a class, schema validation, loop cutoff, retries, fallback, circuit breaker and dead-letter queue. | Errors |
| State and memory | Conversation summaries, tasks that resume, fact memory and prompt cache. | Context |
| Traces and metrics | Per-turn W3C traces with MCP spans, reliability dashboard, alerts and OpenTelemetry export. | Observability |
| A person in the loop | A card for every write, a proposals inbox, approval from messaging apps and undo. | Governance |
| API to integrate it | Agent API with idempotency, waiting for the result, live events and webhooks, and an MCP server with OAuth 2.1. | API |
| Portability | Agents and flows as versioned JSON files, optionally with the text of their documents; signed packages in the Agent store. | Agents as files |
| Choice of model | Five providers and self-hosted models, with provider and model per agent. | Suppliers |
| Enterprise security | The person's permissions, critical actions that do not exist, third-party data as untrusted, budgets and audit. | Security |
No. It is also an agent platform: it can be activated without the business modules and it orchestrates systems that are not part of Fisher Platform through MCP, 28 connectors, any API with OpenAPI, webhooks and its own agent API.
The error goes back to the model with its class (invalid input, no permission, transient failure, timeout, business rejection or internal failure) and a message without secrets. The input is validated against the schema before running. If the same call fails three times in a row, the turn is cut off. Writes are never repeated on their own.
Bastino retries with exponential backoff and honors Retry-After. If configured, it switches to a fallback provider. A provider that fails five times in a row is skipped for a while (circuit breaker). Every attempt is recorded in the trace with its cost.
Long conversations are summarized automatically without splitting tool call and result pairs. An interrupted task resumes at the round where it stopped, without applying again what was already applied. There is also a memory of preferences and of facts confirmed by the person, and knowledge with verified citations.
Each turn is a W3C trace with its model calls, its tools, its MCP calls, its actions and the business record it changed, with duration, tokens, cache and cost. You see it on the Traces screen, export it as JSON or CSV and, if you turn it on, send it to your collector over OpenTelemetry (OTLP/HTTP) with the GenAI conventions.
As a server, it keeps no state: any node behind a load balancer serves any request. It serves lists in pages, caches each person's catalog and has quotas per connection, per client and global. In the published measurements for version 3.82, a tools/list went from 2,052 ms to 529 ms and, with 20 concurrent requests, from 1.6 to 6.8 requests per second. As a client, it defers the tools of large servers: one with 500 tools adds about 150 tokens per round instead of about 51,650.
Yes, with the agent API: POST /api/bastino/agents/{id}/runs with a person's API key. The job runs in the background with that person's permissions. The result can be polled, arrive by webhook, be awaited in the same call (Prefer: wait=N) or be followed live through SSE events. It supports Idempotency-Key. It is off by default and is opened agent by agent; waiting and events are also off by default.
No. Out of the box, background work is run by the scheduled task every minute. If the server allows it (systemd or supervisor), the optional queue worker php spark bastino:worker handles it within seconds: in the published measurements, a job sent through the API starts between 1.6 and 1.8 s after being queued.
If the administrator sets it up that way, yes: each person connects their own account for an MCP server in "My connections" and Bastino uses it in the chat, in their tasks and in the Virtual company. Without an account, it does not offer that system's tools and says where to connect it. By default, the installation's connection is used.
No. Reads are direct, but every write to an external system is proposed on a card with the system, the action and each piece of data, and a person applies it. In addition, each connection decides whether Bastino does not use it, only reads, or reads and proposes writes.
Claude (Anthropic), OpenAI, Azure OpenAI, Google Gemini and OpenAI-compatible servers for your own models, such as Ollama. Each agent can have its own provider and model, with its key.
No, on purpose. An agent always acts as a person or with a person's API key, and can never do more than that person. For an external system, you create a technical user with exactly the right role and its key.
An agent platform is also chosen by its limits. These are the ones in version 3.82.
tools/call. The MCP server answers calls in JSON, which is compliant: no tool emits progress.Link header. YAML: the subset OpenAPI uses (no anchors or aliases).traceparent arriving from outside is stored as a link, not as a parent, on purpose: a third party's id does not decide which trace data is added to. The OpenTelemetry GenAI conventions are still in development.What is not a limit: JSON-RPC batches, the GET stream and MCP sessions. The current specification (2026-07-28) removed them; a compliant server answers one message per request, and that is what Bastino does.
And one more: no model is infallible. That is why critical actions do not exist, protected ones are rejected by the server and every write goes through a person. The model can be wrong; your systems do not change without you.
We show you an agent connected to your tools through MCP, a connector or your OpenAPI, with its traces, its errors and its reliability dashboard.
The screenshots are from versions 3.81 and 3.82 (Workers and My connections) in a demo environment with a fictitious company and people. The turns in the traces were produced by Bastino's real engine with a simulated AI provider (no real provider was called); the MCP servers, the personal accounts, the Jira connection and the carrier's API are fictitious, and the jobs on the Workers screen were handled by a real queue worker. The figures come from the scripts published in the product's technical documentation, measured on the machine described on the dates given. Trademarks belong to their owners; the names of other products only indicate the category it is usually compared with.