The software that manages context, tools and checks around a model.
A harness manages instructions, context, tools, permissions and checks around a model. It creates continuity by retaining and supplying relevant transcript or context on each request, so the model re-reads what it receives. Ordinary inference uses fixed weights, while separate fine-tuning can change them, and provider application programming interfaces may retain transcripts server-side. Read The Harness.
Heard in the wild · Illustrative exchange
How does it know what we agreed yesterday?
The harness saved that record and included or retrieved it for this request.
Give the current task the information it needs.
Context engineering chooses what to load, organise, retrieve and leave out. A useful context window contains the task's instructions and relevant evidence. Read the context engineering guide.
Heard in the wild · Illustrative exchange
Shall I attach the entire deal folder?
Start with the question and retrieve the evidence needed to answer it.
Replace conversation detail with a shorter working record.
Compaction frees space by summarising or otherwise reducing earlier context. It can lose qualifications and exact evidence, so keep durable decisions in memory and prepare a handoff brief.
Heard in the wild · Illustrative exchange
The summary dropped our exception. Now what?
Restore it from the source and record it explicitly before continuing.
A record another session can use to continue the work.
A handoff brief states the goal, decisions, files, verified results, assumptions and next step. It lets a new session recover the task without reconstructing every turn.
Heard in the wild · Illustrative exchange
Can another agent pick this up tomorrow?
Give it the current files, what passed, what failed and the next check.
A reusable package of task instructions.
A skill usually starts with a SKILL.md file and may include scripts, templates and references. Progressive disclosure loads its resources when needed. Browse the public skills.
Heard in the wild · Illustrative exchange
I keep giving the same research brief. Can we save it?
Turn the repeated method and its checks into a skill.
Load detail when the task calls for it.
An agent can inspect a skill's description before reading its full instructions and selected references. This keeps context engineering focused while leaving deeper material available through retrieval.
Heard in the wild · Illustrative exchange
Do all the instructions have to load at startup?
List the available skills first and open the relevant one for the job.
The repository that maps your working repositories and vaults.
In this working setup, root AGENTS.md maps shared repositories and vaults, while home/global.md holds universal rules. Project instruction files add local rules for the agent, which can load a relevant skill.
Heard in the wild · Illustrative exchange
Where should the agent look for the project vault?
Follow the home repo's map and then read the project's local instructions.
A connection to another system's data or actions.
Model Context Protocol, or MCP, is a standard for exposing capabilities to an agent. A connector can expose a tool, with permissions controlling the available reads and writes.
Heard in the wild · Illustrative exchange
Can it read the calendar and change meetings?
Check the connector's granted capabilities separately for reading and editing.
A named line of work in version control.
A branch lets changes develop before they join the main version. Use a separate worktree when agents need isolated files, then review the change through a pull request.
Heard in the wild · Illustrative exchange
Can we try this without changing the main version?
Create a branch and keep the experiment there until review.
A separate working directory for a Git branch.
A Git worktree shares repository history while keeping its checked-out files separate. Give concurrent subagents separate worktrees when they edit different branches, and review each branch before combining changes.
Heard in the wild · Illustrative exchange
Why did the other agent overwrite my file?
Give each editing agent its own worktree and a defined set of files.
A proposal to merge a reviewed change.
A pull request compares a branch with its intended destination and records review and automated checks. State the acceptance criteria and attach evidence from the verification gate.
Heard in the wild · Illustrative exchange
The checks are green. Can this merge?
Read the change and confirm that the evidence covers the requested behaviour.
A user's instruction and the agent's work before it hands control back. A single turn can involve several model requests and tool calls.
A turn usually begins with a user message and may include several inference requests and tool exchanges before the reply. The harness supplies relevant earlier context to each request, creating continuity while inference weights remain fixed. Provider APIs may retain the transcript server-side, and separate fine-tuning can change weights.
Heard in the wild · Illustrative exchange
Was that one reply just one model request?
It may have needed several requests to call tools, inspect results and answer.
A continuing record of work with an agent.
A session groups messages, tool activity and working state under one task or conversation. Its harness decides what history reaches later requests and what survives a restart through memory.
Heard in the wild · Illustrative exchange
Can I close this window and continue later?
Check what the harness saves and leave a handoff brief for the unfinished work.
The model relies on the context supplied for the current request.
Ordinary inference reads the supplied context using fixed weights and does not learn the conversation into those weights. The harness retains and re-sends relevant history, so the model re-reads rather than remembers it between requests. Provider APIs may keep transcripts server-side, while separate fine-tuning can change weights.
Heard in the wild · Illustrative exchange
Why must the old decision appear in the new request?
The model needs that context again, even when the app has kept the transcript.
The system retains information between operations.
A stateful application keeps records that affect later behaviour, such as a transcript, selected files or task status. A harness can provide this continuity around a stateless inference service.
Heard in the wild · Illustrative exchange
Where does the unfinished task list live?
In the application's saved state or files. Check which record the next session will load.
The amount of context a model can handle in one request.
A context window sets a capacity limit under a provider's counting rules. Usable performance also depends on relevance and attention degradation, so compaction alone does not guarantee a good answer.
Heard in the wild · Illustrative exchange
The document fits. Why did it miss the clause?
Capacity tells us what can be supplied. Test whether it can use the relevant passage reliably.
The harness reduces context automatically as a limit approaches.
Autocompact triggers compaction under the application's own policy. Review the retained summary and keep critical evidence in durable memory before continuing.
Heard in the wild · Illustrative exchange
The conversation suddenly became shorter. Did anything disappear?
Inspect the compacted record and restore any missing decision from its source.
A model's structured request to run an available action.
A tool call names a tool and supplies arguments for the harness to validate and execute. The returned tool result can become input to the next model request.
Heard in the wild · Illustrative exchange
Did it actually run the search or just propose one?
Check for the executed call and its returned result in the activity record.
The data or status returned after a tool runs.
A tool result may contain records, files, errors or an execution receipt. Treat external content as evidence to inspect, and verify claimed actions against the tool call and verification gate.
Heard in the wild · Illustrative exchange
The tool returned success. Is the report correct?
Success confirms the operation's status. Open the output and check the figures.
An action the harness makes available to the model.
A tool has an interface that describes its inputs and operation. Its power depends on permissions, and a connector can supply tools for external systems.
Heard in the wild · Illustrative exchange
What can this agent actually do to the workbook?
Inspect the available tools and their write permissions before assigning the task.
A controlled environment that limits what an operation can reach.
A sandbox can restrict files, processes or network access. Its boundary depends on the actual configuration, so check it alongside permissions before running a tool.
Heard in the wild · Illustrative exchange
Can the experiment touch my other folders?
Check the configured filesystem boundary and test the permitted paths.
Instructions supplied in a system role for a model request.
A system prompt can define behaviour, constraints and the task environment under the provider's instruction rules. The harness assembles it alongside other context, and it works with technical permissions.
Heard in the wild · Illustrative exchange
Can one instruction stop every dangerous action?
Use clear instructions and enforce the action boundary through permissions and tools.
Saved information that the system can load into future requests.
Memory may live in notes, summaries, databases or provider-managed records. The harness retrieves or supplies it as context, so check what a new session actually receives.
Heard in the wild · Illustrative exchange
It remembered an old preference. Where was that stored?
Inspect the saved memory or instructions and correct the source record.
A system that uses a model to carry out work through actions and feedback.
An agent combines a model with a harness that manages context, tools and execution. It can inspect results and choose another step within the scope and permissions it has been given.
Heard in the wild · Illustrative exchange
What makes this more than a chat reply?
It can act through tools, inspect what happened and continue towards the assigned outcome.
An explicit allowance to access data or perform an action.
Permissions determine what the system can read, change or execute. Configure them at the relevant tool, connector or sandbox boundary and reserve approvals where required.
Heard in the wild · Illustrative exchange
The brief says publish. Does that grant account access?
The task authorises the work, while the configured account and tool permissions determine whether it can execute it.
Find stored material and bring relevant passages into the current task.
Retrieval can search files, databases or connected services before supplying selected results as context. Inspect source coverage and attribution when using it to reduce hallucination or support context engineering.
Heard in the wild · Illustrative exchange
It found three documents. Is that enough evidence?
Check whether the search covered the relevant source set and whether the passages support the claim.