AI agents now write, review and refactor code in many teams. What was still a single chat window in early experiments is today a mesh of several agents, specialised skills and access to external tools. This is exactly where OpenChamber comes in: the tool sees itself as a control centre for AI agents in programming.
The current version 2.0 brings two innovations that sound technical at first glance but are immediately noticeable in day-to-day project work. Firstly, changes to skills, agents and settings are applied without a restart. Secondly, tool calls can be bundled.
What skills, agents and tool calls mean
In this context, an agent is an AI system that does not complete a task in a single response but over several steps: analysing, planning sub-steps, calling tools, evaluating results, refining.
A skill is a delimited capability given to an agent — for example an instruction on how code reviews should run in a particular project, or which conventions apply to commits. Skills are what adapt a generic model to a concrete project.
A tool call is the moment when the model leaves the pure world of text and executes a function: reading a file, starting a test script, running a search, querying a dataset. Tool calls are the lever with which agents actually make a difference — and at the same time the point where control matters most.
Changing during operation
The fact that changes to skills, agents and settings will in future take effect without a restart seems like a comfort detail. In everyday work, however, this point determines whether an agent configuration is maintained at all.
Anyone who wants to sharpen an instruction and has to stop the entire process to do so tends to postpone the correction — and carries on working with a configuration that does not quite fit. If this hurdle is removed, fine-tuning becomes a normal step during work. This brings the maintenance of skills closer to what teams are already used to from configuration files and linting rules: ongoing, small improvements instead of rare, large overhauls.
Bundled tool calls
The second innovation concerns the tool level. When several calls can be combined instead of processing each one individually in sequence, this changes the flow of an agent session. Fewer individual rounds between model and tool layer generally mean less waiting time and a more compact history that is easier to follow afterwards.
This is relevant for evaluating agent runs: a log of a few thematically related steps can be reviewed. A chain of dozens of individual calls is practically never read through by anyone.
From experiment to controllable process
The real point behind such a version lies less in the individual features than in the direction it indicates. Agentic coding is moving away from experimentation and towards a discipline with its own tools, its own configuration and its own operating model.
This can be seen in the questions that come to the fore:
- Which agents are actually in use in a project, and with which capabilities?
- Which tools are they allowed to call — and which explicitly not?
- Who changes skills, and how is this change documented?
- Can it be reconstructed afterwards how a result came about?
These questions largely correspond to what has long been regulated in classic software development: versioning, access rights, traceability. What is new is that they must now also be answered for the AI layer.
What this means for clients
For companies that buy in development services, the difference between a black box and a configurable system is considerable. A black box delivers a result whose origin nobody can explain. A controllable setup, by contrast, makes visible which rules apply, which tools were involved and where humans intervened.
This is not a purely academic distinction. It touches on topics such as quality assurance, handover to internal teams, handling sensitive data and the question of whether a solution can later be developed further without the original service provider.
It therefore makes sense to clarify from the outset, in projects with AI support, which tasks agents take on, where binding human review steps lie and how results are documented. This clarification takes little time and prevents later discussions about responsibilities instead of content.
Assessment
Tools like OpenChamber shift agentic development from a one-off demonstration to something that can be configured, observed and corrected — and thus embedded in regular project workflows. For web and software projects, this above all means that AI support is increasingly treated as part of the development process and not as a special case alongside it. Anyone commissioning projects today should ask less whether AI is involved and more how its use is controlled and made traceable.
Sources
- OpenChamber: The control centre for AI agents in programming — heise developer News