Agents | Field note

Agents for real workflows

An agent is only interesting when it can survive the boring parts of work: context, repetition, exceptions and handoffs.

The seductive version of agents is autonomy. The useful version is often discipline.

Real work is full of partial context, changing assumptions and responsibilities that cannot be inferred from a topic alone. Building agents for ongoing work and source-based analysis makes those gaps concrete: selecting relevant material, interpreting what changed and preparing the next action are different decisions.

What the work teaches

The strongest agent concepts do not begin with "what can the agent do?" They begin with "where does the work keep losing shape?" That might be between systems, between people, between meetings, or between a decision and the evidence that should support it.

The prototype has to expose the boundary of delegation. What should be automated? What should be prepared? What should be escalated? What should remain stubbornly human?

Reusable lesson

Design the handoff before the autonomy. A good agent leaves a trail: what it used, what it inferred, what it skipped and where human judgment must enter.

That makes the work less spectacular but more usable. The goal is not to remove people from the loop. It is to make the loop worth being in.

Concrete lessons

LESSONS

Shape the workflow

Start with the work before the agent

The strongest agent work does not begin with autonomy. It begins with understanding the workflow: who will use the output, what decisions it should support, what data it can see, what uncertainty matters and where the handoff to a human or team happens.

User research is not separate from agent design

Qualitative interviews with future users can create the foundation before the agent itself is built. Interviews expose the real job: what users trust, what they need explained, what would slow them down and what kind of output a technical team can actually build from.

Design the handoff before the autonomy

A useful handoff names the finding, its evidence, the unresolved question and the next decision. Say whether the agent completed an action, prepared a proposal or stopped for review. Include enough context for the next person or worker to continue without reconstructing the conversation. Human review should have a concrete purpose.

Separate output depth, scope and permission

A briefing and a detailed analysis can use the same method at different depths. Global, sector and domain questions select different evidence. Neither choice should silently grant new permissions. Design these dimensions separately so a request for more detail does not become authority to change the underlying records.

Control the evidence

Keep baseline and live signals visibly separate

For example: a frozen knowledge baseline is not the same thing as a current signal. A live web source can strengthen monitoring, but it should not silently rewrite the canonical picture.

Read-only needs a behavioral and a technical boundary

Define what the agent may read, assess and propose, and what it must leave unchanged. An instruction to remain read-only is a behavioral contract. Check the actual tools and permissions separately, and restrict unnecessary write access where supported. A current assessment can differ from the saved baseline without becoming a register update.

Route to the smallest useful source set

Large agents need source routing. The agent should know which baseline layers, source baskets and overlays matter for the question. Reading everything is not thoroughness; it is often noise wearing a serious coat.

Dates are part of the evidence

For agents working with current signals, recency must be explicit. A source without a date should not carry a time-sensitive conclusion. "Latest" is not a vibe; it is a retrieval and evidence problem.

Keep judgment honest

Classify the source basis before analyzing

For example: full text, partial text, metadata, secondary summary and mismatched source are different evidence situations. An agent should downgrade or stop when the source basis is too weak.

Confidence and overreach must be visible

Agents become dangerous when they sound certain on thin evidence. Good outputs show confidence, overreach risk, alternative readings and what cannot be inferred from the available material.

Do not confuse signal with conclusion

One news item is not automatically a new risk. One biased text is not automatically propaganda. One weak signal is not enough for a strong claim. The agent should compare, qualify and mark uncertainty before it escalates.

Check what a comparison can support

Comparison requires compatible definitions, periods and methods. A changed score can reflect a refined event, a different horizon or revised calibration rather than a changed world. Explain those differences before describing a trend. Comparing source framing is not fact-checking; comparing saved assessments is not an observed time series.

Relevance does not assign responsibility

Role context helps an agent decide what matters to the user. It does not establish ownership, decision rights or a commitment. A relevant issue may require awareness, support or escalation rather than a new task. Assign work only when the source supports a request or commitment; leave ambiguous responsibility open.

Bound the delegation

Skills can become agent modules

A useful pattern is to let the agent own the role, boundary and workflow while a bundled skill owns the reusable method. That makes the system easier to improve: the surface can change without burying the analytical craft.

Test the boundary under real conditions

Test forbidden requests, instructions hidden in sources, missing evidence and unsupported numerical claims. First establish that the intended condition exists in the runtime: an unavailable-source test proves little if the source remains accessible. Record the tested version, observed behavior and unresolved cases. A listed test is not a passed test.

Delegate one bounded decision at a time

Source selection, relevance triage, analysis and proposed updates need different context and authority. Give each worker a clear question and an explicit result contract. A triage decision can select material without assigning work. An analysis can identify a possible action without carrying it out. Use code for mechanical steps that require no judgment.

Calibrate usefulness as well as correctness

An analysis can be traceable and still miss the point. Review omissions, emphasis, role relevance and whether the output helped someone decide or act. Keep user recollection and work context separate from source evidence. Feedback can improve the method without rewriting what the original material established.